Ep 3: Nov 24 – The rise of Agentic AI

I was speculating in October’s newsletter on the continuous development of some brands in the AI chipsets.

• Here we spoke in the first line about Nvidia, which, with sponsorship from big players like Microsoft, Google, and Amazon, was hitting a new peak in the stock market. 
• At the same time, Nvidia announced in October that the new production line for the AI’s most powerful chipset, Blackwall, is already fully booked until the end of 2025, indicating clear gaps between market demand and needs. Interesting to see announcements, for example, from Microsoft, just one week later, on the collaboration with AMD that indeed seems to be entering the competition with Nvidia, and clearly, big cloud providers are interested in reducing the risk of single sourcing.
• Intel announced they will not try to compete with Nvidia’s top range lines with their Gaudi 3 chipset due to the much more advanced status Nvidia is at this time, and so they will focus more on the lower lines competing on price per performance versus the equivalent lower line of Nvidia and other competitors. On top, they showed improvement versus the former fiscal quarter, and that gave positive results on the stock value.
• Qualcomm, on its side, apart from speaking about collaborating with AMD on strengthening its power against Nvidia, has a strong roadmap also in the area of more powerful chipsets for automotive. I will speak about that in the section about cars below. That said, Qualcomm was last month influenced on the stock by the dispute with ARM for the license to use their architecture for chipsets. ARM gave 60 days to solve the dispute or forbid the possibility of producing further chipsets.
• My Thoughts: This clearly impacted the stock of Qualcomm, but my bet, please take as not speculation and not expert, is that the issue will be solved and Qualcomm will get through on the stock value. Reason is how they answered in the communications and the solid roadmap they have behind with many product lines that I expect will bring a solution to the table. On top of that, a dispute not solved on the ARM would deny the production of the chipset for many Android devices within 60 days.
• All these chipset players, but also cloud providers with cloud servers chipsets, computers, and phone companies like Apple with its M and A chipsets, depend strongly on one single player, TSMC, Taiwan headquartered (with all the influences of that with the actual geopolitical situation), due to their advanced production processes. The interesting production plans for TSMC include one more production factory in the US by next year, and with the other factory in development in Dresden (Germany). These will be two interesting hubs for all those future needs and more distributed workloads.
• Interesting side-note that in October, the US government raised the concern that TSMC would have produced on US territory, parts for Huawei, and that would be an infraction of the actual regulations. Interesting to see how this is going to be solved and how it could be influenced by US elections, considering the history of the US and Huawei’s past conversations and restrictions. Clearly, TSMC is working hard to prove they are doing their best to avoid such a case. For example, the recent update of the identification from TSMC on the Huawei case received a lot of attention.
• My Thoughts: Definitely, I still believe, like last month, that some consolidations between AMD, Qualcomm, and Intel could happen, but not clear in the timing and if this will get real. For sure, there are some levels of partnership like Intel with Samsung, in this sense. 
• There is an interesting story with ST Microelectronics that did not have good earnings results, nor has a good outlook for 2025, and is on restructuring, but at the same time seems to have opportunities in the cars market, especially in China, and is curious to see how it will develop further on top with possible AI and health sensors. Interesting that Intel, delegating activities to TSMC in areas where they are missing competencies, is at the same time making wrong steps in communicating about the relationship with TSMC, killing the discount they got for such outsourced activities.
• For sure there is a market opportunity especially in advanced AI chipsets, with a leading player (NVidia) being fully booked for production requests for the next 15 months and not able to take further orders at this time for some specific top lines and a semiconductor player (TSMC) having the most advanced miniaturization and leading on the market but being also a single point of failure. Final note on TSMC, their US production will not run the highest density transistors level that will remain in the leading technology sites in Taiwan for the time being.
• Speaking about AI market forecast, we saw with the end of October and the earnings update from Microsoft, Google, Amazon, and Apple, some interesting trends. Microsoft overachieved the analyst’s plan, but the stock value dropped more than 5% as an indication of concern about the investments in AI not monetizing and possible influence on the Azure future trends. Google, on the other hand, ramped up a little bit, probably as an indication of more confidence, and also related to the former situation, which improved. Amazon was also running well. It seems that investors fear the hype of AI in terms of investments, as it is not recognized by real value generation. This was especially the case of Meta, which was indeed overachieving on earnings but at the same time considered overspending on the AI hype. At the end, Meta, less than 2 years ago, was fully focused on Metaverse (not forgetting the brand name change), spending around 46B$ on that, and the change of focus was highly impacting an important number of employees. It seems quite clear that the AI investments go hand in hand with high value in some businesses, like Pharma, as I also mentioned in my October newsletter. Some more reflections on this are also in the new section named Science and Social below.

In the area of AI, we have many recent updates that are interesting to mention. The hype on AI is still quite high, and there are many conversations that a big part of these initiatives will not survive after the end of the hype in the near future.

• One relevant element for this month in terms of hype, also the title of this newsletter, is the ramp-up of the AI Agents. Now trying to be pragmatic in explanation, the AI Agents look to me interesting as concept that is in principle bringing some level of former robotic automation augmented with GenAI but then augmented with more autonomous capability, moving from a static, rules based approach, typical from an RPA to a behavior approach based on description of the purpose of the agent making it able to adapt to the requests and also able to execute more autonomous activities potentially fully e2e and in some way take decisions for some business processes. It’s interesting in this sense that the announcement made by Marc Benioff, CEO of Salesforce, about its AgentForce in October at Fortune, after his initial announcement in September, was again on Fortune. There is clearly some pressure to deliver AI in the evolution of Sales tools in terms of real value add, and in September, Marc was speaking much about the possible future trend to have AI agents to operate on behalf of sales employees, introducing even the concept of a license not anymore per user but per AI agent. 
• In October, he moved more in the direction of attacking Microsoft for its Copilot, as a side effect, in my opinion, due to the announcement just a few days before from Microsoft on the integration of Copilot GenAI capabilities in its Dynamics 365, competing in the same segment. An advantage I see here is that on one side, Salesforce has strong data from its customers and long experience in this field, and that could build proper AI Agents on top for automating, improving sales decisions, where Microsoft could deliver that experience across multiple components of their ecosystem, up to data layers and office, but maybe would require more work to integrate.
• Parallel, ServiceNow, playing more and more as a platform of platforms, pushed also strongly in the direction of AI Agents that could allow for some pieces of processes to be automated and integrated with GenAI. For sure, that could be interesting in areas where some business processes have already been implemented or are planned to be implemented in such platforms, and make some automation. I’m also thinking of a typical case, like growing an automatic knowledge base of an enterprise through AI Agents, just to mention one. 
• As well, SAP with its Joule is playing in the area of AI Agents automation, with an interesting hype on procurement processes, which seems to me quite beneficial as an improvement to prioritize.
• Another case, quite obvious, is linked to software coding AI agents no longer assisting but mainstream in code development.
• The proof of the pudding will be in how much these agents will be able to reduce repetitive activities while interacting with GenAI capabilities and how well they will be able to participate in the process of autonomous decision-making, based on the training, to accept or deny in the decision-making process. For example, there are activities around having some processes of insurance managed by AI Agents that, in some cases, would work to make decisions but would have a second-level human approval. Also, activities in healthcare with NVIDIA in terms of AI Agents to execute avatar-driven hospitality. On the other hand, it will also be sensible to set boundaries where these agents will not be effective. 
• I’m reflecting, and here is one of my correlations, on Apple’s research that is also recently progressing on the limits of the mathematical capabilities of LLM in terms of interpretation. I’m thinking in this sense how important it will be to set limits within which the AI Agent will be able to make decisions with mathematical interpretation through GenAI, initially restricting areas of application and progressively evolving as the models develop further. Just to make a brutal basic example, no one would trust to have an healthcare AI agent to provide recommendations on pill doses based on a GenAI conversation of symptoms that could have a mathematical bias and indicate wrong quantities.
• I see some good business cases, for example, using AI Agents to automatically feed the internal knowledge base, evaluating which content would be interesting to be aggregated. Here, different players would have different advantages to make this effective as one of the core capabilities of an enterprise. Another one would be customer services, sales ordering, procurement, as already mentioned, hiring processes, and possible data consolidation. 
• When I look at the bare IT support processes, an acceleration about AI agents is already happening for years, first a few years ago with chatbots, then last year with automatic ticket GenAI summarizing, and automatic knowledge base, and now progressing in automatic ticket interpretation and process automation.
• One more reflection here on the AI Agents, but here my thinking is more blurred, would be to use such agents not only to make full automated business processes, but also to build and maintain updated and consistent interfaces between applications implementing a business process and data consolidation across structured data, including data cleansing activities. In this sense, I see benefits in more repetitive tasks as mentioned before, and as well on the data lake/data mesh level, for example, to help to identify and expose relevant data to be made available for publishing. 
• Where I see another area of interest and is tight with the part of interfaces/integration is in the area of IoT, but for this, I have a new separate section below. 
• The other area where I would see AI Agents is in the HRIS hiring platforms and job applications platforms, and you can find more described in the section below on HR Platforms. There are already aspects of AI content moderation, for example, in platforms replacing employees for their activities. You can find more details in the section Science and Social, where I elaborate on the effects of AI automation and the job market on a specific case in social network platforms.
• Continuing on activities automation, an interesting recent update is from Anthropic, a famous Open-AI competitor that just announced a way to make many activities happen behind the screen of a computer, automated through an AI recognition engine with GenAI. In this sense, Microsoft also played with its highly controversial Recall, going in that direction even if it started more from a time-machine approach. What we can see is that many administrative activities done by employees on a computer could be automated, especially if they are repetitive, but not only if they follow certain paths. I can imagine this will allow us to improve process mining activities for all those areas that today are not fully managed in ERP and are human intervention-based but would also play a risky role in monitoring activities executed by employees on computers. So the opportunity and the privacy risks will have to be managed accurately and seem clear to me that the edge of application will be quite tight.
• Also important for the AI Hype will remain the aspects of intellectual property. I mentioned in the October Newsletter on the risk of AI Mad due to training AI with data coming from AI. In this sense, the effort from DeepMind is to set watermarks in text to identify when AI was generated. As we progress, I believe there will be more and more techniques to watermark and to go around intellectual property contents, and that will impact the AI training and capabilities we will see developing, and the consequent reliability on bias in decisions.

• It was interesting last month to see aspects related to the power of computation benefits where, thanks to the advances in GPU capabilities, a researcher, a former Nvidia employee, worked out to identifya new biggest prime number, 16 million digits bigger than the former one, part of the 52 known Mersenne prime numbers. This was somehow also driven by the big hype of AI, which developed the GPU market to the next level.
• Also in October theNobel Prize for Chemistry was won by the co-founder and the Director of DeepMind, thanks to the software with AI capabilities built that is able to predict the 3D structure of proteins from their amino acid sequences. This shows how much the AI and, especially, Pharma are in general tightly connected in the development of the most advanced area of AI capabilities that can bring real value. Similar capabilities develop in the area of new materials and will generate other value streams in Chemicals and R&D of other industries. There are, on the other side, other activities in the Healthcare and AI where the GenAI could bring value, but also fear based on job cuts, and bias in evaluations. Here, the transformation is happening too fast and without proper attention, which would seriously impact the reputation of those firms not properly built and trained their AI platforms.

In the Social, we saw:

• It seems that the AI has already generated some aspects of redundancy in low-mature markets, where content moderation has already moved to an AI-managed social network platform, TikTok.
• There is a big hype coming now on the remote work reduction strategies, hiding a silent firing strategy driven by AI automation introduction. There is a clear sensibility on how this will evolve, and just last month, I was speculating on the too-aggressive back-to-work with zero remote strategies, especially reflecting on the new GenZ attitude. That could fire back as businesses will lose efficiency if the AI automation does not bring practical benefits at the level of scale expected. At the end, the workforce has more flexibility to adjust their activities than a vertical AI agent, so the risk of anticipating reductions based on AI hype will cause, from my point of view, a risk of losing flexibility versus future needs. I think it will be crucial to make a proper process, transitioning to new employees’ roles while AI will take on more repetitive activities and adjust organizations sizing according.
• In the environment, instead, we saw:
• Interesting that I referred last month to the growth of electrical consumption from Microsoft due to AI and the wish to restart some old nuclear plants. At the same time, also pushed by the impossibility to wait for long decisional processes from the government, Google and Amazon are also pushing in the direction of their own nuclear energy production, based on small nuclear plants near the datacenters, also allowing less impact and dependency on the whole grid. It’s coming to me a few questions, one for example, how this will be secure, also from a military point of view. If before the datacenters would be sensible targets to attack data, computing power, and connectivity to companies, this is going to introduce a new possible sensibility type, even if these are mini-nuclear reactors. It’s also interesting for me to understand if this verticalization of the market, where an IT company starts to produce its own energy, will bring, later on, the other part of the equation, the fact that these companies will also start to influence the energy production and delivery market.
• However, out of all this, it remains the core question that I raised in August. What will happen with datacenters in those countries that decided to go against nuclear power? For example, Germany and what about those countries like Italy, where Microsoft announced last month that they will invest 4.2B$ to build datacenters in the next two years, considering that they will have no nuclear plants possibilities? Will this impact the overall energy capabilities, or will it impact the cost of the services that, if hosted in those countries, will have a completely different cost? Will this push the change of strategy of some countries in terms of nuclear usage? Will this slow down the evolution of Clouds and AI in those countries?
• One of the challenges in the datacenters grows due to the fact that many of the AI activities require strong interconnection, so they are not planned to be executed on geographic distribution, limiting the max capacity of computation by plant. Some algorithms could be built to be fully distributed at a certain stage, or for example, the phase of training of AI could be distributed, and that would reduce the physical restrictions, allowing to run computation across geographic datacenters, but it seems still in an early stage of development and not necessarily feasible on a large scale of algorithms. Definitely could have a strong influence on the overall strategy of datacenters’ AI capabilities delivery, and is an element I would keep an eye on.

In the IoT world, there has been an interesting update coming last month. 

• On one side, the 5G seems not to have taken much on the market in terms of usage versus investments from telco, and some speak about the 6G that is yet to be on the market, but an interesting element that is raising is about the eSIM rise for IoT. This is currently existing in around 30% of IoT devices, but there is a clear trend that is going to rise heavily in 2025, also driven by the opportunity to make such connections more reliable, and behind all this could also change the complexity to have different devices having proper connections and integration to specific clouds to feed data fabrics with less edge complexity. Here I would like to remind you that 80% of the data of a company is NOT the bare ERP data, and the OT represents a big portion of that 80%. So relevant is to consider that a Data strategy, designing how to feed data into one or more concentrators and transform, has to also design a proper correlated AI concept. 
• I strongly think that an AI strategy can’t decouple from a clear Data strategy as a base. So in this sense, as I was mentioning in the AI section, there is a big chance of transformation to come to the IoT world in terms of automation of processing and transformation with GenAI capabilities. Here, I think it will be valuable to have eSim capabilities to make it more flexible to maintain an updated and in sync wide variety of IoT equipment, and the AI Agents could help to make a good first consolidation/clean on data at the edge level (so speaking more on edge computing in this sense).

However, having 70% of equipment not yet capable of working with eSim will require, most probably, a lifecycle of replacement that will continue over the next 5 years. So in this sense, I’m not sure how quickly the IoT part will evolve worldwide with new capabilities, but definitely interesting to keep an eye on and consider linked to the strategy of OT standardization, also from a security point of view.

In the electrical we have seen that the announcements from Elon on his expected volumes for next year caused a +20% in a few hours in the stock of Tesla and BYD, which is making a quite consistent roadmap evolution of its segment and earnings. 

• Limiting to what is relevant on technology trends change, an interesting announcement from Qualcomm goes in the direction of predicting a bigger volume of technology hardware for cars coming in the future. Their roadmap is filled with many chipsets with low consumption and higher performance to allow for maintaining several screens and sensors in a car, also with limited cloud connectivity as edge equipment, feeding a central unit of transformation with much data. This would allow for generating stronger, smarter content and capabilities in the cars, like sensing by the feeling of people and their speech, the need to change some setup (i.e., someone feeling cold, saying in the car to someone, and getting automatically his/her section of the car with temperature adjustment automatically). 
• The car segment is under stagnation in Europe, also due to the entry cost of cars, especially in the electric segment, and it is interesting to see what will be the relevant customer trends in this sense. There was less than 6 months ago an analysis reporting that a certain number of people felt distracted by so much technology and screens, so the future in this sense has to be adjusted, and the profound market stress in this area could bring some changes. 
• Anyway, looking at the announcements, seems that ST Microelectronics would benefit from China’s needs in terms of semiconductors in the automotive so also interesting to keep an eye on the market evolution coming especially from Asia.

Right in my former Newsletter of October, I was referring to the absence of any quantum computing Center of excellence in Europe, as also Draghi was mentioning in his report, considering that out of 10, 5 were in US and 4 in China.

• Announcement in October that in Germany, IBM will start the first Quantum Computing datacenter. I think this will be important for European competitiveness, but will require bringing innovation and technology into the hands of the European lead rather than being a US-hosted and owned-only innovation. 

Quantum computing is relevant today for many research areas, especially in materials and molecules analysis and simulations. It will get more hype as it will combine more with AI. I generated for that sense a free channel speaking about Machine Learning and Quantum Computing named MacQuantum that I update only on specific ad hoc cases, contents I think can be relevant.

Over the last few months, we saw the ramp-up of automated AI solutions with GenAI. In the October newsletter, I mentioned the future trends related to Human Digital Twins and the progressive combination of human personalized content aggregating personal data related to health, well-being, and work. You can see more details in the former newsletter.

Looking more at the specific aspects linked to the job market, there is, from my perspective, a ramp-up in the generation of personalized job applications, including per customer application automatic cover letters. 

• There are several platforms already on the market that are getting more capable of integrating with platforms for job posting and are able to personalize with GenAI to fulfill the full application process, including cover letters. Some of these solutions aim to achieve even more than 750 applications submitted per day. Some real users’ experiences speak about automating around 20 applications per hour, reaching a total of nearly 3000 applications automatically applied, including GenAI-generated cover letters personalized for each firm and role. This is for me bringing an obvious new element that even if enterprises have already implemented an advanced HRIS Hiring engine with AI, there will be a progressively more AI to AI triggered interface where the AI solutions to find jobs, will hook how to rate their paying users, to make them succeeding in passing the first filter of job application. Unless properly organized and filtered, there is a big risk from my perspective, to stress the HR organizations of firms to manage the wave of applications that could make a match. This will ramp up, again from my perspective, as Gen Z will be more on the market and will look to apply to jobs in an easier way from their perspective. There is a common opinion that GenZ has less patience, also driven by the flow of data they manage and the time to react they are used to, and this could make them quite willing to find easy ways to access the market, rather than through a selective research process, mass-applying through AI-supported processes. 
• I can speculate for example that some enterprises could decide to make the first application process shorter having many hundreds or thousands applications received in just few hours or even minutes instead of days and the job automated application engines on the market would hype to identify and apply in realtime as one job would be on the market available with continuous opportunities monitoring, leaving all those people not using such engines, simply too slow to react to the market offers.
• Again, I would believe some AI Agents could be the next level of automation of filtering and evaluation on both sides, from job application platforms and from hiring platforms. What is clear is that these are going to be more and more AI-to-AI interactions, as I can see, and each enterprise will have to find itself prepared to manage such a process change.

In this section, I try this month only to remind you that 

• In October, the NIS2 regulations went live, and it seems that 34% EMEA companies didn’t plan accordingly and shifted budget from other areas to the NIS2 activities.
• Finally, as a last reflection of this month, looking at those figures, out of 34% of companies that missed a proper budget in the EMEA for NIS2, the budget was retrieved for 30% from recruitment, meaning impacting either other areas to innovate, or IT Ops. 

When I think that Cybersecurity is a well structured department from years in enterprises and can still struggle to get properly funded, it makes even more tangible my speculation of last month in the October newsletter when I was assuming that, if not properly set boundaries around deliverable, benefits, ROI and risks, AI will risk to take progressively budget from the IT Ops budget and other IT innovation initiatives budgets, impacting on the overall IT roadmap.