Guide to AI Safety and Compliance for Startups with Fion Lee-Madan

If your use case require you to use predictive AI, traditional AI, so existing platforms like Snowflake, Databricks, DataIQ allowed a non-specialist to easily feed in a dataset to train a predictive AI model. So that's their goal of democratizing building and development of AI in the predictive AI world. If you're talking about generative AI use case, then AWS Bedrock has free developer training for their AI tools and products. So does Microsoft. So I would suggest that's where you start, taking some of these free resources from the big techs, because ultimately they want you to use only AI so they can sell more GPUs and AI products and services. My co-founder and I actually met at Founder Institute. In the Kitchener-Waterloo cohort back in 2020, right before the pandemic. We have two in-person sessions, and after that, everything was locked down and everything was moved virtually. The idea of FairlyAI came from my co-founder, David Van Bruyn. He previously had another AI startup where he was using NLP, natural language processing, to do cyberbullying detection on social media. It's during that experience where he realized when you are building AI models, and back in 2014, he was building AI models from scratch when there was no foundation models yet. He realized quickly that to build AI models is one thing, but to test them is a completely different beast. Hence, after that company got acquired, he wanted to start Fairly AI. Myself, my background is in computer science, with an MBA, grew up in the e-commerce platform era. By 2018 is when a lot of my clients were moving off of my enterprise e-commerce solution, which got acquired by Oracle for a lot of money in 2010. And then, you know, everyone was moving to Shopify essentially. So I was looking for something new to do, and that's why I joined Founder Institute. And met David there. So FairlyAI, we are providing a governance, risk, and compliance solution to help provide quality assurance for AI, specifically for 3 criteria: performance, risk, and compliance. And part of today's topic is to, you know, do our introduction, but also I personally believe in, you know, understanding history is the best way of going forward in a lot of cases. So what are some of the challenges and case studies that have already been established in the past about the adoption of AI or building AI startups or building AI models? And then how do we solve them? And then I know a lot of you are just starting out, so having free tools that you can use right now without having to pay money definitely is going to be very helpful. So my goal is to help you, you know, get started in your compliance, safe, and reliable AI adoption journey. So I talk about David, my co-founder. He actually also taught ethics and advanced math at Cornell and Berkeley before his last startup. So math and ethics essentially is the core pillar of what we call responsible AI today, and so we are ahead of the curve. I talk about myself already. Our CTO came from IBM, IBM Research. Her superpower is productizing IBM Research, which is a super useful skill to have when the technological landscape is moving so fast today. Hasan Patel is our Director of Global AI Policy Compliance Engineering. That translates to a lawyer that can code. Also very super useful skill to have with this very fast-changing regulatory environment. So we did Founder Institute. We also did a whole bunch of accelerators after that, including Accenture's FinTech Innovation Lab, where we actually built our first prototype with 14 tier 1 banks' risk and compliance teams. And then we did Techstars in Toronto, won a whole bunch of awards backed by some government and also nonprofit organizations. Technology partners include Microsoft, which I'll talk about why you should pay attention to that as an early-stage startup in the AI world. So some of our selected partners, because of our expertise and we started early in this space 4 years ago, so some of the more forward-thinking companies like EY and BCG actually were our co-design partners in the early days. Standards councils like the British Standards Institute, ISO standards, and the Councils of Canada also worked with us to pilot a few of the ISO standards for AI risk management. I mentioned about Microsoft, Snowflake, and SAS, also one of our, some of our selected partners. So moving on to challenges and case studies. So for those of you who are new or may not have built any AI yet, right? So what is the main difference? Why all of a sudden there's this seem to be like flurry of regulations around AI? So traditional software, 1 1 2, just like when you use your calculator, is deterministic. QA gives a very clear pass or fail status when you're trying to QA traditional software versus AI, artificial intelligence models or application. 1 1 usually equals 2 is probabilistic, right? When you use ChatGPT, I think now everyone finally understood, thanks to ChatGPT, that you don't always get the correct answers, right? And what that means when you're trying to QA such a probabilistic product application is that you are literally making performance and risk tradeoffs. And who should be making these tradeoffs? What are the right tradeoffs? That's the problems a lot of people who have been doing AI a long time are trying to solve today, including us. So what are the risk tradeoffs, right? So there are many different ways of categorizing risk. So this is just our approach based on our customers' and our partners' feedback. There's operational risk, which is really based on the inherent risk of your AI application, such as where you are going to operate, the jurisdiction, which sector you're building the model for, right? Financial services, you know, healthcare, some of these are very highly regulated industries. So the inherent risk of those AI model applications can be higher. And then it's the actual use case, methodology and where you are in your project life cycles. Technology risk inside of financial services is actually pretty well understood. They call it model risk, and there's a whole bunch of companies that are already helping financial services companies to deal with model risk, model risk management. The specific framework is called SR 11-7 in the States, It came out of the last financial crisis back in 2011. And a lot of people credited why the financial industry didn't collapse during COVID is because they have this model risk management practice in place for more than 10 years. So I'll speak a little bit more about that in the solution size. And then compliance risk with the increasing regulatory challenges in the AI space. Compliance risk becomes a huge issue, especially in a regulated industry as well. So what do these categorical risks translate to? So when practically speaking, when we talk to our customers, which is a lot in, I would say 70% right now is in financial services, they talk about security as like one of the top priorities because I think it is one of the most well-understood risk, type of risk, traditional software, you have cybersecurity risk. Now LLMs have increased the attack surface of these kind of risks. So security risk is definitely top of mind. So if you're selling into enterprise company, you already have to do SOC 2 compliance, for example. Now you may have to worry about how to prove, how do you prove that you've done testing for your AI security risk. Privacy, same issue as traditional software. Now you need to think about Your training dataset, does it have privacy, private data? Is it gonna leak private data? Financial risk, right? If your AI model deals with, you know, profit and loss type of situation, that financial risk is definitely top of mind for you or your clients. Then it's actually hallucination. You would think that hallucination comes in number 1. Actually, from our experience, like actually, falls a little bit behind security, privacy, and financial risk. Then environmental is becoming a hot topic. There have been research papers that published some of these large language models uses a lot of energy when you're training them. Copyright IP is also another issue that has been popping up. OpenAI is being sued with multiple lawsuits on their copyright issues. If you're just starting out to train your AI model, that's something you can correct now. Versus like after you train your model and then you go back and try to fix your data source. It does not work that way. Ethical issues then come, kind of became like a non-starter in today's day and age. And I'll show you some real-life examples in the case study section. And then when you combine all of these things together, reputational harm is one of those things that, again, if you address it right from the start, you don't have to worry about it. Then some of the more prominent AI researchers, AI academics have talked about extreme risks of AI. While this is a hotly debated topic, people can use AI for weapons, weaponization, job losses, etc., etc. We're not going to focus on that today. Like I said, it is a hotly debated area. My personal view is AI is bringing a lot of goods to the world. My personal example that I love to use is education. Personalized education using AI is huge. I live in 5 different countries in the last 15 years. One of those countries is Malaysia, Kuala Lumpur. They have a lot of refugees from Syria. When you're a volunteer teacher at these refugee camps trying to you know, teach kids that are from a different background, from different grades, you know, may not even speak English, right? As a volunteer teacher, you really are very limited without technology, without tools like AI to help you, you know, do instant translation, do personalized learning when, you know, you could be teaching anyone from kindergarten to high school. So with that said, I just want to put it out there that I am actually very pro-AI, but We just want to all do it safely and reliably, and also making sure it's compliant. So jumping into case studies, what can we learn from history? So this is actually back in 2018. Some of you may think AI is new. It is not new, definitely. So Amazon had an AI hiring algorithm that was basically preferring men for all tech jobs. In the current day and age, if you're building any decision-making tools that uses AI and you don't do bias testing, that is already a failure. Google had a photo recognition algorithm using AI that was, you know, recognizing Black people as gorillas. This again was back in 2018, 2017. We actually talked to the team that did this, Ultimately, it's a data issue. For some of you that are also new to AI, 70% of AI problems came from the training dataset, garbage in, garbage out. That's why data labeling, after all these two episodes that were pretty well known in the AI ethics circle, data representation, having the diverse set of training data, has become a huge issue and is one of the core principles in a lot of the responsible AI frameworks today. But you would think that solved the problem, it won't happen again. So, Google Gemini recently rolled out a few months ago. What happened was they now have, you know, corrected the diversity and inclusion problem, but then forgot to test for historical facts. So, they had this example where you were using Google Gemini to generate pictures of American founding fathers and they included people that are indigenous, Black people, which historically is not correct. They overcorrected one problem and caused issue in another. This is not an uncommon issue in the world of building AI. For those of you who've done AI longer, you probably know where I'm going. I'll keep that in mind. A different kind of risk, different kind of problem. Lawyer cited nonexistent case loss. That's hallucination. I think for a lot of you that have been using ChatGPT or pay attention to your news, you would know that ChatGPT doesn't always give you the correct answer. Sometimes it hallucinates, gives you information that does not exist. Once you know about that, of course, you would make sure that you properly fact-check anything that ChatGPT tells you as a lawyer especially. However, this happened in May 2023. Another lawyer had the same issue in February 2024. So why, why did this lawyer not know that the other case already existed and not be so stupid and did the same thing using ChatGPT and cited a non-existent case, right? So how do people not know to learn from history. Here's actually one of my explanations. What you read in the news, like Google News feed, your Facebook chat, your LinkedIn chat, is all already driven by AI algorithm. When I look at my news feed, it's all about these AI accidents, so to speak. But if I look at my husband's news feed, it's very different. It's all about trading stocks because that's what he does all day long. Your worldview of AI is very different from mine. I guarantee you that if you're not in the bubble that I live in, when you just go on Google News. That's part of the problem we're trying to solve in some of the free tools I want to give you so that you now can become part of my world of looking at what are the existing history of AI accidents, so don't repeat the same mistakes. Here's another one which you may or may not have seen in your newsfeed. Air Canada had a chatbot that was lying. It gave out false pricing information. So that's another type of risk, is really if your risk and compliance team, your policy team, doesn't understand these things, right? In this case, the Air Canada legal team actually, you know, went to court with this small claim, which ended up to be an $800 fine because they were found guilty of, like, providing false information and be held responsible for it. But for whatever reason, their legal compliance team said instead of like just, you know, settle it, refund the $800, they actually went to court. It became huge international news on CNN, on Forbes, on BBC, right? It became a huge reputational risk, reputational harm for Air Canada. So that's another type of risk, the reputational risk, even though, you know, the actual financial loss is relatively small. The latest one is Scarlett Johansson versus OpenAI. If you had followed some of the news or in my bubbles, you would have seen this news that the voice assistant from ChatGPT called Skye sounds a lot very similar to Scarlett Johansson, and OpenAI decided to take it down right after they released it. Copyright issue, that's definitely a huge topic right now in the legal circle. So when you're starting out as a startup founder, that's something you can control still and you want to control right now so you don't run into the issue down the road. Because once your AI is trained with these copyright data, you cannot get rid of them and you basically have to discontinue your model, just like OpenAI have to discontinue this AI assistant called Skye. Back to my very first slide, you are dealing with complex performance and risk tradeoffs. Another analogy that we like to use is like the game whack-a-mole. You are trying to optimize for accuracy score perhaps, but you may suffer in other areas such as fairness bias score, or it could be anything like historical facts. So how do you solve these risk trade-offs? Who should be solving these trade-offs? Who should decide what risk trade-offs to solve? That's the next section of my presentation. But I want to present another view for some of you that are from a technical background. Traditional software testing is very serial. You start with unit testing, which is a lot of your performance testing, and then you go into integration and system testing, which is where a lot of your risk testing comes in. And then you go into compliance testing, which is where a lot of your user acceptance testing comes in. If you do it serially, back to the whack-a-mole game, you might solve one thing but break the other thing. So our approach and methodology is you need to do everything at once. Plus there's global AI regulations coming, EU AI Act. For some of you who are planning to sell to Europe or exist in Europe, you probably are very familiar with that already. Outside of EU, if you are in Canada, US, you know, India, any of these countries and planning to expand to EU, you also need to pay attention. Not only because, you know, there's like you're being governed by the EU AI Act, which already has passed. Some of them will be enforced by end of this year. Some of the rules will be enforced by 2026. But it is also considered the gold standard, just like the GDPR, the Data Privacy Protection Act that came out of EU. It became the gold standard of data privacy act around the world. Everyone's looking at the EU AI Act as the gold standard for AI regulations around the world. So if you're paying attention to that, you're probably already become ready, business ready for any of these new AI regulations that's coming very fast and furious. If you're selling to existing regulated industry, then, you know, it's even a non-starter. You already have a lot of existing compliance requirements. Doesn't matter if it's an AI system or not. I'm gonna skip these examples in the interest of time. So if there's one thing you need to remember, back to, What is AI? AIs are probabilistic models. This is a famous quote from a British statistician, George Box. All models are wrong, but some are useful. This is actually the core tenet of model risk management in financial services. So the headline is what people like to use, but the actual quote is what's in the picture, and I personally like it better. Remember that all models are wrong. The practical question is, how wrong do they have to be to not to be useful? Okay, think about that for a second, because what can go wrong? So maybe you're building an AI to do marketing productivity, but there are AI models, models that are being used in airlines. Boeing's 737 MAX 8 disaster was a modeling error. This is one of the examples Professor John Hill from NYU loves to use. He teaches model risk management at the graduate level, and it is one of our technical advisors. So it can kill people. Another one that I like to bring up is, you know, using AI in medical decision-making situations like organ transplant. If you have had family member like myself who have gone through, you know, organ transplant and have to be on the waitlist, these AI algorithm are literally dictating who gets to get that organ, like, you know, down to the decimal when you have this concept of, you know, categorizing the data, labeling the data, the training dataset that you're using to train these AI algorithms. So think about that. While your use case may not be, you know, as high risk or high impact, they are definitely the reason why the regulations are moving very quickly because they are truly high-risk use cases that are being deployed yesterday. So how to solve? The first thing you can do in this day and age is adopt a responsible AI framework. I just put one up here, Microsoft Responsible AI Principles. I like to bring that up because they actually have been doing responsible AI the longest, and we know that team well. Some of their, you know, responsible AI team member has moved on to other organizations like consulting companies, BCG, Montreal AI Ethics Institute. Their training at Microsoft actually have been spread to other organizations and it has been proven in many cases that it does help as a type of risk mitigation. I have a list in a later slide that I can show you. But that's the minimum you can do in this day and age. But why? Why even do it? If you take a step back, how do you test an AI? It is very much like testing a car. You cannot test every single component of a Toyota. You cannot test every single car that's going out to production. So how come Toyota is so reliable? It is all about the failure. What and how to test, and what do you do after the defects are found, right? You think it's like thousands of tests they need to run, but no, in the lab before it goes to production is 6 stringent tests. We call this stress testing, scenario testing, and like making sure every car can run 200,000 trips, making sure the power window can work for 10 years, 1.5 million windshield wiper wipes. That's stress testing. Scenario testing, extreme heat and cold. And then what do you do once you found defects? Like, you know, zero engine part defects. Once you find one defect in a lot, the entire lot is scrapped. Same as the bis— consistence. Same thing at IBM, uh, CTO came from IBM, you would think everything needs to be 100%. No, it just needs to be 100% test cases attempted. If it's a performance issue, 95% success rate is okay. But if it's a highly regulated industry in a high-risk problem like cybersecurity risk, cybersecurity defect is zero tolerance. With that said, then the this very serial process becomes— this is our view— you can do everything at once, combining all those 3 concepts together: stress test, scenario test, regression test, backtesting. For performance, risk, and compliance, every time you're upgrading LLM, you know, from Llama 2 to Llama 3, or you want to try Google Gemini, you need to test everything again. Ultimately, you're left with what we call, the industry risk experts call, risk residual risk. Just like a car, doesn't matter how many times Toyota test it, you still need a seatbelt. So in the AI world, what are some of the seatbelts, the mitigations the more mature companies have already put in place? You probably have heard of the word guardrails. So code-level guardrails probably is something that you're most familiar with. But the more mature the company, they actually have policy guardrails. And as a startup, I highly encourage you to talk to that with your legal team and lawyers, counsels, right? So having the right terms of use to cover certain use cases that, you know, if someone use your chatbot, for example, that's outside of the purpose of your chatbot, then you're not liable as a policy guardrail. And then just like cybersecurity risk, there are cybersecurity insurance you can buy. There are now AI insurance you can buy, and we have some partners that offer that. And from our point of view, well, combining that all together, if you're building in a high-risk or high-impact use case, That's when you need a GRC system like ours that does policy management to have policies and controls, test case management to make sure you have the right test cases in place, and actually doing the testing to continuously measure, monitor, and manage your AI risk and compliance. So free tools. Follow us on LinkedIn or take a snapshot here, and then I'll quickly jump into the list of free tools. I know we're— maybe we'll spend 5 minutes on this, Emily, and then we can wrap up and do a Q&A. So if you go to our website, fairly.ai, under resources, you guys actually inspire me. I just added this link today called free tools for startup. Okay, and my plan is to continue to update this list. Some of the accidents, if you don't have a Google News feed like my bubbles, you go to the AI incident database, is funded by the World Economic Forum and another nonprofit called Partnership on AI. You can pull up a list of accidents that has already been logged and there has been over, I believe, 2,500 accidents that have been logged. You can search it by different company, especially for the ones that You already know some of these big tech companies or smaller companies may be your competitors, or you can search their name to see if they have some AI accident that has already been locked. Don't repeat the same mistake. Back to that lawyer use case, make sure you don't do stupid things that has already been reported by news agencies. Then the next part is if you're not sure what regulation you want to go into, we actually have a map of global AI regulations. If you click on the top right corner in our blog, you actually get an enlarged map with filters. So you can filter by countries, you can filter by industries and sectors. So, and it's color-coded, you know, so the legend is on our website, you can read that. Green and blue are the ones you definitely need to pay attention to because that means it's already passed and it could be already in enforcement. Then the next thing, for those of you who doesn't know what Responsible AI framework is, we actually have a list also. You don't have to use the Microsoft Responsible AI framework if you don't want to. The NIST AI Risk Management Framework, if you're operating in the United States, is actually a front-runner. If you're trying to look into certification, the ISO 42001, which is listed somewhere down here, is another one. That large organizations are also starting to adhere to and paying money like Responsible AI Institute or ourselves to help them be compliant to these international standards. And then for those of you who are operating in the EU or just want to get a sense of what the EU AI Act is, the IAPP, I forgot what it stands for. I think it's like International Association of Privacy Professionals. I have a very handy matrix that you can use to determine what— which part of the EU AI Act applies to you, depending on which type of operators you are. Are you a provider? Are you a deployer? Etc., etc. And then for those of you who want to dive deeper into that, the EU AI Act Committee itself actually has an EU AI Act compliance checker already that you can use. To fill out a questionnaire to determine which part you actually definitely need to adhere to. For those of you who may not have raised your first round of funding yet, depend on a lot of free credit, or maybe you have raised money, you also love free money like ourselves, Microsoft actually have a Microsoft for Startups program called Microsoft for Startups Founders Hub. It gives a lot of Microsoft Azure credit. Here it says up to $150,000. There's actually another tier that we're on called Accelerated Pathway. It gave us $350,000 in Azure credits with Tier 4 support, which means we have one-on-one technical support from the, from the Microsoft teams. But you have to grow into that. So start with this level, right? It's easy to apply with your LinkedIn profile. But with those Azure credit, not only you can build your SaaS platform on it, but They actually allow you to use it on the OpenAI Azure Studio, which is good because it's very hard to get a GPU. In many cases, you get on the priority list with this Azure Founders Hub. Also, they can increase your token quota limits, etc. There's a lot of benefits if you're thinking of building AI. What else can I show you? Responsible AI toolkits. Some of the examples I show you, John Hills, Professor John Hills' History of Model Risk Management lectures, Montreal AI Ethics Institute is a great resource. A lot of these organizations I listed are, you know, free for anyone to join, including startups, but are backed by big tech. That's why I listed here in particular, like the Partnership on AI Responsible Practices for Synthetic Media. Is formed by the who's who in the big tech world. But if you're trying to use AI to do any type of synthetic media, image, video, then you may want to already look at the framework that these big techs are already using, and it's free for you to participate as well, same as ML Commons. I'll stop there and take Q&A.

Thank you. Thanks, you all. That was great, really a lot of information to digest. And you have this great reaction, people are happy. And for those who were asking about the presentation and about the video, yes, you will receive the video, and if you're able to share also the deck with us, that would be even better, and we'll send it in the same email. So, uh, yes, it was really useful, especially the free tools. It was a lot there to remember. So let's go to the Q&A now. Uh, everyone feel free to add more questions. We have a couple here, so I'll just We can start with the first one, written speech. As new models come out, they have new licenses associated with them. Having legal teams review them can get expensive. What do you suggest a startup can do to understand the usage terms of open-weight models? A question for Garima.

Thank you for that question. From our experience, if you join an accelerator like Founder Institute or, you know, something that's further along like Techstars, usually there's free legal resources as part of the mentorship resources provided by these accelerators. So that would be my recommendation if you're looking for free legal advice as an early-stage startup. So, another approach to that is really partnership with university. So, in Canada, we actually have a lot of grants partnership partnering with research institutes. So, in your country, you might want to look into those opportunities where, you know, the researchers also have a lot of experience in dealing with especially open-source licensing models.

Thanks for this advice. And next question: do you use tools like Unity Catalog for governance?

I actually do not know what Unity Catalog is, so the, so the answer is no.

Maybe whoever asked this question, uh, her name is, uh, anyways Please, you can clarify. And which AI insurance do you recommend?

Okay, so I know Munich Re, based in Germany, does offer AI insurance. Happy to connect you, drop me on LinkedIn, we can chat about that.

Thanks for that. And yes, you can connect What are some strategies? Open source foundation models after they are found.

Yeah, so that is majority of where our customers are dealing with and even ourselves dealing with because we do use AI in our own product, right? So depending on the experience of your team, Some of your team members may be more experienced in dealing with these defects in models, and they would know to, you know, try to find another foundation model that performs better for that specific type of defects. But ultimately, back to the performance and risk trade-off, you have to find a model that works best for your use cases. For some of the companies that are, you know, more resource-constrained, they rely on external partners like consulting companies You have more money, you go to the Big Four. You have less money, you go to smaller consulting company. For people that want self-service tool, that's when they come to us.

Great. Next question about hallucination. How is the hallucination tackled?

This is a very hotly researched topic. Depending on the use case, but if you're talking about using LLM for knowledge base, no Q&A, type of tool, so having a RAG system, Retrieval Augmentation Generation system in place with a factor database so that you can query the factor database for source of truth, ground truth, combine that with LLM has been the main methodology to deal with hallucination lately.

Okay, hopefully that helps. Thanks. We have time for maybe 2 more questions. Next one is from Jetal Kedia. What are the steps to incorporate AI in startups?

There has been a lot of tools that has democratized the building and development of AI. Depending on who you talk to, some of the data scientists we work with, right, are using the term traditional predictive AI versus, you know, new generative AI. So if your use case require you to use predictive AI, traditional AI, so existing platforms like Snowflake, Databricks, DataIQ allowed a non-specialist to easily feed in a dataset to train a predictive AI model. That's their goal of democratizing building and development of AI in the predictive AI world. If you're talking about generative AI use case, then AWS Bedrock has free developer training for their AI tools and products. So does Microsoft. I would suggest that's where you start, taking some of these free resources from the big techs, because ultimately they want you to use only AI so they can sell more. GPUs and AI products and services.

And Jethro also added to his question, what could be the challenges for doing so, for incorporating AI, except finances, the financial part?

A lot of the risks I talked about, that's definitely one of the challenges. Figuring out, you know, A, the ROI of the AI use case, And then is the risk and reward worth the ROI? Because if you're trying to build an AI use case as a startup in a highly regulated industry, you definitely need to already pay attention to these compliance requirements, which is already a challenge even before AI existed for a lot of fintechs, right? Trying to make sure they have SOC 2 compliance in place when you're trying to sell to big banks, which costs money, right? Happy to chat more about that. Yeah, drop me a LinkedIn.

Right. And our last question before the networking, I will put this on stage. Question is by Rafael from Berlin regarding IP rights when applying to accelerators or incubators. Are there any benchmarks?

In my experience, accelerator incubators doesn't take any of your IP rights at all. They helped to help you to figure out what the right IP strategies are. So I wouldn't worry about that. Where we came into IP rights issue is when we're working with research partners. Some university requires to share the IP, some do not.

Thanks, thanks a lot for answering this last question. And before we jump to our networking lounge, would you like to share Any parting words? Anything else you want to add to your presentation before we go?

Yeah, so follow us on LinkedIn and also would love for you to join our waitlist. I think on our website there's a get started for free button. Honestly, I'm already scheduled onboarding in August right now, so I mean, depending on where you are, you may or may not need our tools yet, but happy for you to join our waitlist and be, you know, kept in the loop for when you're ready.

Yes, don't forget to connect to Fionn Lee and check out Fairly, get more information there. We will send you the deck and the video. And also, if you want to learn more about FI, we will post now in the chat a link with all the information you need will also be on a table in the networking lounge. So see you there, and thanks a lot again for joining us. Have a good day, or have a good night. See you soon.

See you.