In this issue

1. The reckoning is approaching
2. Santa Claus exists
3. Cooling is another problem to solve
4. Something to know: the black box paradox
5. Let's talk about AI as it deserves
6. Look how powerful my AI is
7. An AI tool to try
8. Our project to bring AI into marketing
9. Copilot news! The AI agents control room is here
10. Controls in AI projects: leading the way

 

 

1. The reckoning is approaching

 

We know the story of Amazon, which operated at a loss for years before becoming a money-making machine. So discovering that OpenAI doesn't generate any profit makes us think it will in the future. But exactly how?

 

Today ChatGPT is ubiquitous (even ​Boris Johnson loves Chat Chippy Tea!​), however OpenAI spends on computing power and energy much more than it earns, also because the majority of chats come from free users. To give some numbers, current costs are estimated ​around $20 billion per year and annual investments estimated at $175 billion, compared to the current (presumed) $13 billion annual revenue​.

 

The initial hope was to force users to become paying customers, but the sudden arrival of alternative AI models did not allow this strategy. Also because the challenge is aimed at a company like Google which, unlike OpenAI, has its own cloud, processors, models (Gemini, now more performant), as well as other sources of revenue. OpenAI is seeking alternative business models, but at the same time Sam Altman has publicly declared that he has ​activated the "red code"​ within the company to focus resources on improving ChatGPT and maintaining leadership in user numbers. Three years ago, it was Google declaring "​A New Chat Bot Is a 'Code Red' for Google's Search Business​", but times seem to have changed. Also because the barriers to switching from ChatGPT to Gemini or another AI system are really few. Indeed, we would all be happy to abandon an unfortunate name like "chat gippity"! 

 

For the record: Amazon did not generate profit for years because it spent heavily on investments, but its daily business was healthy and generated margin. For OpenAI, however, every single ChatGPT response generates a loss. Personal prediction: the "red code" is yet another marketing gimmick and in a month OpenAI will say "we rolled up our sleeves, here's the new ChatGPT that's fantastic!".

 

 

 

2. Santa Claus exists

 

This morning I woke up with the urgency to ask Copilot if Santa Claus exists (see screenshot below). The answer (fortunately!) was affirmative. The chatbot also knows my previous conversations and since I am a Pro user, I imagine it knows my age as well. Yet, when it comes to Santa Claus, it is ready to declare his existence.

 

It certainly pleased me to think of the child who, on their dad's tablet, asks ChatGPT and gets reassuring answers about the effectiveness of their December letter, but at the same time it makes us reflect on how AI can be "modeled" not to tell the truth.

 

As reported ​by (the always excellent) Post​, several studies are analyzing the impact that chatbots could have on upcoming elections, considering the increasing tendency of people to consult them on various topics. Research so far agrees that AI is perceived as "sincere" and that it is ​capable of changing opinions and voting intentions​ of those who use it. Once upon a time, people said "​it's true, I read it in the newspaper​", "I heard it on TV", "I saw it on Facebook", and now – perhaps – "AI told me so".

 

The problem of verifying sources remains the same, but I don't hide that AI seems more insidious in not openly declaring whether it has a biased position or not on controversial topics. Or rather, chatbot responses on debated topics usually do not take a clear stance, explicitly stating the lack of universal consensus, yet if they managed to make it say that Santa Claus exists, how difficult would it be to train it to confirm the goodness of Trump's policies?

 

 

 

3. Cooling is another problem to solve

 

AI requires energy to perform large amounts of calculations, but also to cool the computers that carry out this activity. Even experiments on new quantum computers highlight the need to lower their temperature. This is why there is great interest in technological innovations that promise new ways to "cool".

One possibility studied (by Microsoft, Google, the Chinese Highlander, and others) has been ​to place data centers underwater​, in an environment naturally very cold and relatively close to the population. The results are encouraging and research shows that already at 40 meters below sea level, energy savings exceed 30%.

Starting from this idea, other hostile environments for humans but well-cooled, such as the Moon, are being considered. It is precisely our small satellite that is being targeted, for example by the American startup ​Interlune​, which aims to collect large quantities of helium on the Moon (very rare on Earth and used in cooling systems), or by ​China and Russia for the construction of a nuclear plant​.

We haven't finished exploiting our planet yet, and we're already organizing for the next one? Fortunately, the crème de la crème of Italian science ​is studying how to build factories in space​, in a responsible and sustainable manner.

 

 

 

4. Something to know: the black box paradox

 

A black box is the image usually given to a system that receives inputs and returns outputs, but where it is impossible to look inside to understand its functioning.

 

In the field of AI, the paradox is that on one hand, we want to know the mechanisms by which AI models work; on the other hand, this transparency is only possible when computers perform simple tasks (as ​pointed out by Prof. Rawashdeh of the University of Michigan​). The mechanism by which ​a software determines if a food is or is not a hot-dog​ is easy to explain. Conversely, ​an app that recognizes any food and provides its caloric content​ works thanks to a complex system that is still difficult to explain. If the app makes a mistake, it is not easy to understand why.

 

Let’s transfer this reasoning to contexts where the choice becomes very important, such as choosing a therapy, recognizing a pedestrian, or evaluating solvency for granting loans or mortgages; in these cases, AI is required to be "​explainable​", i.e., to show the steps and reasons behind the provided answer. However, these are precisely the complex cases where transparency becomes difficult to achieve: thus we must choose between a powerful AI we cannot trust and a reliable AI that is not powerful. 

 

Emerging techniques exist to visualize how a model makes decisions, which sometimes risk becoming "explanations for humans," not necessarily true. There is therefore the danger of building dashboards, graphs, and heatmaps that give a sense of control but are actually just post-hoc narratives, convenient simplifications for managers, regulators, and public opinion.

 

 

 

5. Let's talk about AI as it deserves

 

At five years old, my son broke his leg while skiing. After a month at home, the kindergarten opposed his return to class because he was still using crutches. I went to the school office and started yelling, like a madman. I never yell, but in that case, I decided to adopt this communication mode as a strategy to obtain what I believed was a right. In general, we adapt the way we speak depending on the situation we find ourselves in.

 

This seems to happen also in conversations with AI chatbots for assistance and complaint management, to which ​people increasingly speak angrily​ because they now know that the AI model will recognize a greater sense of urgency and speed up the handling of the case. Already with Alexa ​it was demonstrated​ that although the system was designed to understand our language, we speak to it with very short, loud, and simplified phrases. The ​research​  also tells us that when an AI system pretends to be a person, we are more irritable and susceptible.

 

 

 

 

6. Look how powerful my AI is 

 

Anthropic ​has just announced​ that it blocked a cyber-espionage campaign, which occurred in September 2025, likely sponsored by China (which denied it). The interesting element is that the attacks were carried out with more than 80% of the activities performed by Claude Code, an AI model by Anthropic. We must consider that some of the targets, about thirty organizations among big tech, finance, chemistry, and US government agencies, seem to have been hit.

 

We thus discover that just as many activities are being automated in the "mature" companies, the same is happening in the dark side of the force. Even the "bad guys" are using AI, for example, to perform work that would once have been done by a human hacker. Which is not surprising.

 

Some independent researchers ​have raised many doubts​ about the actual autonomous capability of the AI system to perform the activities presented in Anthropic's report, which in practice has been accused of exaggerating the case, as proof of Claude Code's power and thus advertising itself. For those leading an organization, the real question is not whether to believe Anthropic's story, distinguishing what is already technically possible from what is mostly storytelling, but rather understanding how to update their defense model both against external AI systems trying to attack us, and ​to control the behavior of internal AI agents​, which act with permissions we ourselves have provided.

 

 

 

 

7. An AI tool to try 

 

 

We know that "prompts" are the requests we send to AI chats like ChatGPT, Gemini, Copilot, etc. For those working in graphics, a good prompt really makes a difference and obtaining beautiful images with AI is all about the prompt: you need to know the styles, imagine what you want to achieve, and describe it accurately. A not-so-simple activity that requires professionalism. Precisely for this reason, a company proposes itself as a prompt marketplace.
​Promptbase​ is a real marketplace where we can see the final result obtained with a certain prompt and then buy it for a few dollars, so we can use it to recreate the images we are interested in, but with the style produced thanks to a specific prompt. If there is someone with a need and people who can meet that need, I find it brilliant for someone to position themselves between these two subjects by creating a marketplace from which they can earn small commissions on the commercial exchange.

The downside I see is that there are already many sites that provide ​incredible prompts for free​, or offer a kind of "reverse engineering," thanks to which starting from an image you get the prompt useful to recreate it. As an example, I took the (very elegant) image below and uploaded it to ​one​ of these sites. I thus obtained a prompt, which I then modified to use the original style and represent a charming university professor of almost 50 years. I would say the experiment was successful, although it doesn't resemble me.

 

 

 

 

8. Our project to bring AI into marketing

 

 

A large international organization in the food sector had a clear need: speed up the production of multilingual content while ensuring consistency, quality, and reuse of already available information. The manual creation of articles, social posts, press materials, and informational documents required time, coordination, and continuous exchanges between internal and external teams. To address this challenge, AGIC created an integrated platform that combines collaboration, automation, and generative artificial intelligence.

At the center of the project is Azure OpenAI, which allows generating texts on demand: the user selects the type of document and language, attaches any reference materials, and the tool produces the requested article - or translates it - maintaining style and consistency with the provided content. From the same input, different formats can be obtained, such as Facebook posts, Instagram stories or press kits, optimizing communication across multiple channels.

Supporting this, SharePoint Online ensures a structured and shared document management, Power Pages offers intuitive interfaces for the presentation and review of content by internal and external users, while Power Automate orchestrates the request, translation, and approval processes, reducing time and minimizing manual interventions. The result is an intelligent and scalable tool that allows the client to accelerate the production of global content and improve the efficiency of the entire editorial cycle.

 

 

 

When we have a beautiful project, but I can't tell you the client's name, I feel like this.

 

 

9.Copilot news! The AI agents control room is here

 

 

In the last month, at AGIC we met two companies with the same request: "we now have various AI solutions internally, we need a control room to monitor and manage them". Done and done. Three weeks ago Microsoft launched ​Agent 365​, which we could consider a Control Room, dedicated to the company's fleet of AI agents. Thanks to a centralized dashboard with telemetry, dashboards, and alerts, IT managers can monitor the agents' activities in real-time, measure their performance, verify how they interact with users and data, evaluate the return on investments, and be ready to intervene in case of suspicious behavior.

 

Agent 365 also allows integrating agents with productivity applications, data, and workflows already in use (Word, Excel, Teams, SharePoint, Dynamics 365, and others). On the security front, the platform extends the same protections that a company applies to human users to AI agents: with tools like Microsoft Defender and Purview, it is possible to detect and block threats, monitor data exposure risks, apply controls, and retain detailed logs to ensure compliance and auditability. Here ​you can find some remarkable videos​ showing Agent 365 in action.

 

 

 

 

10. Controls in AI projects: leading the way

 

 

Many companies are investing in artificial intelligence, but ​according to a Bain study​, AI projects often stall when control functions come into play: legal, risk, compliance, procurement, finance, HR. Apparently, the problem is not the presence of controls per se, but the fact that they are involved too late, only to give final approval. Under these conditions, with little transparency on use cases and no shared criteria for evaluating risks and benefits, the response tends to be defensive, with delays, endless requests for further analysis, or a simple "no".

 

I am reminded of the ​2022 Italian F1 Grand Prix​ at Monza, when the Safety Car intervened a few laps before the end and there was no time to restart the race, with the cars crossing the finish line behind it. Everything was in order, but zero spectacle. Similarly, the corporate control system slows down AI projects, and only something very banal is obtained from them. Bain's study argues that organizations capable of successfully introducing AI into their processes transform control functions from "gatekeepers" to project stakeholders involved from the ideation phases of use cases, to define acceptable risk thresholds and minimum common criteria, as well as leverage existing risk assessment frameworks.

 

 

About me

Hello, I'm Francesco Costantino, university professor and Director of Innovation at AGIC. Passionate about technological innovations and a firm believer in a future better than the past, I enjoy sharing and experimenting with new AI tools available, as well as observing and reflecting on digital evolution.