Introduction: Trusting AI When the Stakes Are High

Moving to the first questions that we want to ask you guys. We want to ask your opinion on this.

Scenario: AI Predicts Your Best Client Will Churn

AI predicts your best clients will leave tomorrow.

So let's picture this scenario, right? So to make sure you're still alive and with us. But they seem happy today. What do you do?

You have three options here. Trust the AI and look for a replacement. Ignore the data and trust human intuition. or wait and see what happens to avoid risks and i would love to hear lorenzo what would you think

Answer: Why Predictive Models Often Beat Intuition

about it that's funny because we do a lot of anti -churn prediction models you know so models that will predict the probability the likelihood of a customer abandoning the company and churning so since i do this as a job i will trust the ai and look for a replacement

basement but you know if we think about predictive AI you know predictive AI is able to read throughout all of the data that you have about a customer so that means that your you know view of a customer will be always limited to the thing that you know about the customer but if you have a lot of data so I think that the point the main point is data if you have a lot of data about the

customer and the AI leveraging this data will predict that this customer will leave, most likely the AI will be right, because they will go through transactions and orders and invoices and claims and customer service calls and whatever, and have a comprehensive view of the client and a predictive view of the client of the probability of churning.

The “Butterfly Effect” in AI Adoption

100 % so I have a question for you before I was taking notes and thinking about the butterfly effect this event it's called butterfly effect for a reason a small change in the day as you said as more change in data quality or productive predictive flow can actually start an amazing domino effects in the business right so and also on productivity and return of investment so um

One Small Change That Creates Big Operational Impact

from your perspective what is the small butterfly like the the first business process or operational touch point that can optimize the predictive AI creates the most significant domino effect like if you could choose just one small change that can create that domino effect what would it be?

yeah um so again it's funny because in terms of butterfly effect before a gentleman from the audience said that maybe you have the war and the world will will affect the prices of your raw material and the price of products and blah blah so um again i think that the thing is is focusing on the core operation of the company.

So if you need to do a small change to have a big impact, for example, in the profit and loss of the company, you need to go through the core operation.

So, for example, if you are a manufacturing company, you have to go to your factories. If you are a CPG company, you have to go to your products and stuff like that.

and you need to rewire the whole process around AI and this you know by implementing this you can do a small change and you will have a huge impact afterwards.

Sure thank you I definitely agree with this any thoughts from the audience so remember that you are part of this roundtable yeah any comment on this, you can go. Yeah sure.

Lorenzo, I would ask you to repeat the questions for the sake of the recording.

From Research Idea to First Customer

So he was asking how I discovered that my idea was worthy of becoming a company I think and so actually that's a pretty funny

story because I was doing my PhD at the University and I wrote to a blackboard some pipelines of processing and the whole idea of you know taking this data having like at the beginning that it was having a model that will be able to generate a model like a predictive model you know and i wrote down this blackboard into my

professor studio and after a couple of hours a manager of a huge company coming to his office to talk about something and he looked at the blackboard and said oh that's really interesting what is this and the professor said wait a minute i will call lorenzo we will start talking and And the manager was so interested that he said something like, if you do this, I'm going to buy it.

And we started with the first client before even starting, like, really developing the product. You know, we have only some ideas in the back end. And after that, so this customer is still our customer.

it's Teleperformance which is the largest company in the world about call centers and we started working with them and after them we made deals with other

call centers and then with banks and blah blah so my advice is always to start as soon as possible to sell actually sell like signing contracts your idea if you are talking about b2b before even wasting you know a lot of time in developing something that maybe it's not interesting thank you probably

AI in Culture and Tourism: When Storytelling Meets Truth

Scenario: A Museum AI Guide Invents a Beloved Anecdote

we'll move forward to another question for you guys so let's picture this scenario a museum AI guide invents a fake historical anecdote that tourists absolutely love. What do you do?

Let it slide. Entertainment is what counts.

Block its historical truth is untouchable.

Fix the AI to include only verified facts or charge an extra ticket for this specific tour. Let's see what you think.

Design Choice: Verified Sources, Clear Boundaries, and User Trust

For sure, this is a thing, right? So before Ricardo was speaking i was thinking that museums but other uh you know structures needs to make sure that the the information they provide are verified right so our questions for ricardo is also

i was thinking that there are re already several ai assistance for tourism culture and public services right so including institutional ones such as um i think about julia by roma capitale

So, what really differentiates LotzArt from the solutions that already exist?

First of all, I really enjoyed the last option and I created an extra ticket.

But just if you say it's a fantasy, it's a fiction guide. it's a fiction guide now of course in fact as I said before we cannot create a fake historical because what we do is to create all the storytelling all the interactive games based on the source, validated source by museums, by cultural institutions something like this

so we had the one of the first events was at the Capitol Museum and I had a very good feedback from a lady that told me I was very skeptical about this kind of experience I told it was like yes some fake some very banal thing just to have a show and it was so it was very impressive because because he knows that it was true, what she heard, but also in a very easy way, interactive way.

About the other question, Julia is a very interesting tool, of course, but it's not vertical. It's more like ChatGPT or Cloud or something like this. so what we want to do with with the RNG is to block to yes to create to not not make the eye on the World Wide Web so not not all the information he can he it can achieve but also some some information that we validated and so uh it's more reliable for this and more easy to also to validate for the users sure any comments and questions from

Why Not Use General Chatbots? The Case for Vertical Assistants

the audience on that because why were you speaking i was thinking also you mentioned Chachapiti, for example, so why should a museum, an event such as this one, or like a city, choose your product instead of simply using Chachapiti, Gemini, or Claude, right? Because I think most of the people in this audience already use these tools, so what really changes with lots of art.

Yeah for this I want to speak not more about culture but for example for events as I say before when you have an event like this one or I also had a PhD and so I attended some political sciences and not so hard like him and I I don't tell the story about the board.

Now, I went to this conference, attended this conference, and it was really difficult to understand what is the speaker that want to hear about it,

what is the team that will talk about it, they want to go, where is, at what time.

MSOEV, this program, I don't know if you have the whole experience, a very huge brochure with a lot of pages, and you don't know, you're like,

okay, maybe if it's a PDF, you can Ctrl -F and then search your speaker and whatever you want.

But it's very difficult to attend in this case. and in CGPT or Gemini or cloud you will never have this kind of information so we want to have just this kind of information and in various way like it

now with conversation we are very able to interact so like I can say I want to speak I want to listen Omar okay so what's up we will tell me what what time it will be what is the team will face and etc etc and so in a very easy way in a very interactive way you will have all your information on your WhatsApp.

Thank you Riccardo.

Now last questions for you before having other questions for our audience. Let's see what you think. I'd love to hear from the speakers.

Autonomous AI Agents: Profit, Control, and “Scheming” Risks

So AI predicts your best client will leave tomorrow. AI agents work alone overnight and makes you money but violates company guidelines.

Reaction. Shock reaction. Rewarded in businesses results matter.

Turn it off. Company rules are sacred or reprogram and restrict its its autonomy. Blame the engineer who set up the software.

Misaligned Subgoals and the Need to Monitor Intermediate Steps

That's a good question. This is one of the real problems we are facing, not that AI makes money for us, that's a dream. The problem is the so -called scheming.

You give a task to the AI, the final task. So please finish this software, I want you to achieve 95 % of completion. And the AI puts its head on this target. But to achieve this target, the AI fixes some middle targets that maybe are different from the final one.

So the truth is that we don't have to intercept this kind of things. We have to intercept the middle tasks that sometimes, it's strange, but are against the final one.

AI has to complete my software at 95%, but I need to make a system upgrade, so I have to switch off the machines because I need to, I don't know, to change something, whatever you want. If the AI understands this, sees the fact that I want to power off the machines as a

problem and maybe starts and it's a matter of fact there is a lot of literature starts making fake logs faking sentences in order to avoid the power of the machines so the real point is that this is a dream but when you let this kind of agents work for several days you have to think these kind of middle tasks that are a little bit different from your final one otherwise they can compromise the

result okay so it's something that can happen okay for sure so and I also was

Ethics and Governance: Minimum Safeguards from Day One

interested by your recent TEDx in Athens right because you talked about a critical point in which is in the butterfly effect really important which is we need ethics and ethics needs humanity right so in a market that

pushes companies to integrate AI quickly to stay competitive, what are the minimum practical ethical safeguards that a company must establish from day one to ensure the domino effects of the butterfly effect? Is there any minimal safeguard that companies should take to guarantee ethical approaches

approaches to the adoption of AI?

Privacy, Transparency, and Responsibility as the Baseline

And that's another very good question, because in Europe we have the rules, we have the Act, even if the Act is going to be postponed, etc. But my personal feeling is that this is not enough.

It's not enough because Europe is Europe, after that we have the US, we have China, and everybody has different, a little bit different rules.

So for sure, the real point here is not the dream, you are asking me a lot of questions about dreams. The real point is to find a common set of rules, maybe a medium set of rules that could

be followed by everybody, because for sure, for example, for example, Lorenzo was talking about having data in so -called sandboxes because otherwise banks never allow him to work on the data.

okay privacy is the first problem okay but it's not the only problem because you know we make a lot of events on ethics because the other problem is that sometimes people trust too much about ai so also transparency is a huge problem the transparency of the the entire process and the expandability of the predictive models okay it's not a problem so at least the privacy Transparency and responsibility are, I believe, the basement.

Any comment from the audience?

Audience Debate: Regulation vs Innovation and Competitiveness

I have one. Yes, please. I don't think so.

I think that there are a lot of laws and regulations already about credit scoring and privacy and anything from the GDPR to the AI Act. And they're all in Europe.

But when we think about how the American economy is going and how the European economy is going, we need to understand that the more laws you put in, the more companies will struggle into implementing innovation. So when you do that, you are protecting citizens, yeah, right?

right but then you don't you know you you can't complain that our salaries are the same as 20 years ago because if you allow companies to experiment freely as in the u .s you will not come up with you know a dictatorship but with a system of free markets that will be able to in which the company that is the most advanced and the most fast in implementing this technology will win.

So about ethical AI, I don't think about it much as a problem, but more as a European problem, type of dream.

I know that we have a lack of time here, so I need to rush a little bit more. I'm sorry to just jump to the Q &A question because I'm sure that you have some thank you so we move forward to the next slide please thank you so any questions so far you can either raise your hand or enter it in the system if you're shy and you don't want to raise your hand but there are non -shy people

people here so there is a question over that yeah this is another problem you're trying to regulate something that is still in shaping, that is still evolving, and it's very difficult to do that. And hopefully, from my point of view, technology will be faster than regulation and will bring us to the next industrial revolution, faster than bureaucrats.

but you know to if you want to do that you can do it you know for example the GDPR was a really really good example of regulation because it was not about the technology it was about the principles of you know privacy and data protection so no matter the technology but the principles

needs to stay the ai act is much worse if you read it have a lot of technological things in it so it's not working because in the meanwhile that they were writing it agent has come in you know so the things that they are trying to do in restricting our technology and not you know having some values and principles that every kind of technology needs to follow although it's really, really difficult for regulators to try to protect people from evil use of this technology if they only focus on technology and not on principles.

I don't know if you guys want to add something. I believe that from my ethical point of view, as I said at the beginning, we need a very small set of rules, but we need a set of rules.

for example AI act is a messy but in Europe we cannot have we don't we cannot have now the social scoring applications that's a matter of fact maybe we are behind but it's a warranty for us from a technical point of view you are right

it's a race and I don't know what is gonna happen for example the new the new version of QN, the QN4 version is very clear fact now. Chinese company cannot even media GPUs but they use that Huawei Ascend 950 and QN4 is a extremely good model so anyway they overcome the ban and I'm working and even developing new ways of working better with worse GPUs so So the race is a matter of fact.

The rules are breakable. But sometimes breaking rules leads to better solutions. Because I don't know if you checked the paper about the new coin version. They made really a huge work with GPUs that are new GPUs from Huawei, but are seven, eight times less powerful than NVIDIA ones. So they made an extreme effort on software to more or less reach the other mainstream models. If you check the benchmarks, they are very close to them now. So maybe this kind of ban can be also good.

Other questions from the audience?

Risk, Investment, and Learning Through Failure

I think the most problem is the regulation. situation, we think it's cashless, so we don't have to pay as European Union. What we need now is investments, but not like they are in this situation.

We need what I think is risk I think we have some capital went to that hold some risk bank that holds some risk institution that also risk and we have to accept also that companies can fail because what we don't accept especially especially in Italy I think

is nobody has to success nobody has to fail so we are in the average what we we need, I think, is to fail, also to learn what went wrong.

I'm in the creative and cultural industry, and I know a lot of companies, they are in the same situation. We are, yes, in the plateau, so it's very difficult to improve, but because we don't see the failure also in the same time.

I think we have to accept the risk, especially in this situation in which U .S. and China but also Ukraine now, also Israel, go fast because they accept the risk. They accept the failure and it's not cashless, this kind of results. But we have to accept this. That's my opinion.

Thank you. Thank you.

Keeping AI Costs Under Control in Practice

And talking about cash, a question for Rosario. How do you contain costs? Okay.

Containing costs is one of the reasons why we started working with these autonomous agents. For a big enterprise, it's totally different because they can spend, they recharge the cost to the customers. When we started using the autonomous agents that write code for us, the problem was exactly this one. the cost can explode because they use a lot of tokens so what we do now but we are a small lab

so maybe it's applicable to labs that our dimension is working only with subscriptions we don't work with api because as i told you the the very beginning okay we save the operational state in this case on some databases but there are also some commercial uh tools for example temporal

we reach the token limit five hours which token limit the system waits five hours and starts again starting from the state is saved on the db on or otherwise on the on the framework for managing the state it starts exactly from the same point because we save the state doesn't have to recover the state to use token to recover the state again so wait five hours and restart and because these

These agents can run for hours, for days. We don't matter so much if they stop every five hours, okay, and restart. So from our point of view, we just work with subscriptions.

The cost is subscriptions, anthropic 200 euros per month, and the same with others. But anyway, there is a cap, a limit on the cost. Clear.

Where Things Go Wrong: Limitations and Failure Modes

so burning tokens it's an issue for most of the people using AI at the moment so 100 % not having that problem it's a win situation for people so any constraints any you know dark aspects of running such a project so has been any talking about a failure any you know moments where something went differently than as you expected during the process so now you

talked about us you talked up to us about the process and the idea so during this process has there any moment where you actually had any you know no moment

Oversight, Testing, and the Value of a Detailed Plan

Yes, for sure, because in this moment, maybe the strongest tools, I'm talking about the open source side, are Automaker, Autoforge, and another one named Aperant.

And it's okay, they can run for days, for weeks, but the software, they provide you at the end, especially if you don't oversight, I don't want to say constantly, but time to time, okay. Okay, the final software still needs some testing, some debugging.

But anyway, even now that they are not perfect, maybe they can do 70%, 75 % of the job for you. So, at least according to my experience, it's very important to prepare a very good, a very detailed plan at the beginning.

So, before we were talking about costs. If we have to spend money, spend tokens on having a good plan. Make a very good plan with a very good model, oboes. After that, when you start implementing the margotasks, maybe you switch to Sonnet.

But time to time, at least now, you have to double -check time to time. But you don't need to constantly double -check. I'm quite sure that next month, this kind of framework tools can manage end -to -end. But it's my opinion.

Thank you. No, no, for sure.

Human Skills That Make AI Work

And I would love our speakers also to share some of the darkest moments because we have talked about AI, how can AI can be actually used in organizations and in processes. But of course, there is an issue in trust the AI, trust the results, not be able to really comprehend AI. And there are sure some struggles that startups such as yours face during the process. And that might make it more human to the people that want to adopt AI, right?

So I also have a question for the three of you. So is there any, let's see, human skill or soft skill that you think can really help the organization or people to understand and adopt AI? Any answers is fine from anyone. Because humans are in the middle of this change.

We talk about how massive AI is, but we forget often that AI should be like something that helps us to go faster, but also go smarter and also be able to do the repetitive task and do the different things that we mentioned before. but in this process the human being is a key part of the succession and achieving of the goals, right? What do you think?

Critical Thinking, Creativity, and Better Prompting

I think that the most important soft skill is the so -called critical thinking, because AI receives orders and if we are lucky, executes orders.

This means that, previously we talked about the contest window, The context window can be as big as you want, but it's a technical, statistical, mathematical context window.

We give orders to AI, and AI maybe makes our orders be done. The AI doesn't execute our desires. So if the orders are not well -formed, we don't think critically on what we are ordering, we can receive different outcomes. So my point of view is that the critical thing is the most important thing.

One skill for each of you, so we go to the last question. Yeah, I totally agree. And I think also human creativity as a critical thinking is crucial.

And for example, yes, in our workflow, for example, for multilingual production, we start from a human actor so from human actor we can replicate in different language but we start from human and then it has to be an accelerator from different things and so yes I totally agree with Rosario and

yeah critical thinking because you have to put the prompt so it's up to you what but also they generate.

Can I answer this question? For sure, which is perfectly on point, right? Yeah.

I'll read it for the audience.

Building Trust in Data: Explaining Predictions to Decision-Makers

When your predictive models show a completely unexpected trend that contradicts a manager's gut feeling, how do you help them trust, we spoke about trust before, their data over their own experience without creating a friction?

This is a really good question. Yeah, that's a really good one. It happens every day.

So, you know, from my point of view, I blame it on the AI. I don't know, ask Eli, you know.

Explainability and A/B Testing as a Bridge to Adoption

But basically, the key point here is having explainability of the predictions. For us, it's crucial. You cannot say to an oil company like ENI that the price of oil is going up 3 % next week without having an explanation for that, because they will not take action if they don't have an explanation for that.

that means that we only use explainable models or if you if you say that customer is going to churn or a customer is not going to pay its debt you need to explain why you think that you know why the model predicts it like that so you really have to show the data the proof and obviously Obviously the choice, the final choice will always be the manager.

So we do a lot of A -B testing for that. And maybe you will try your strategy and then you will try Eli's strategy and let's see which one performs the most.

Yeah. 100%.

Conclusion

So if there are no other questions, I would really love to thank our speakers to have brought...

Thank you. Thank you. Thank you, everyone.