Ep 11: Jun 25 – The rise of Agentic AI – Chapter 2 – The Orchestra

• The chipsets war, as I often referred to it in the past newsletters, remains a key topic to monitor the market’s evolution from my point of view. The AI supremacy run is relevant because who is achieving first an AI able to optimize itself across different areas will gain a competitive advantage over the others and will keep the advantage over and over, making it quite difficult for competitors to catch up.
• The big tech are clearly focusing on optimizing themselves as their priority, involving a big resources optimization on the AI especially in areas that can be automated. I will speak more in detail in the section on AI and Workforce Transformation about how this is accelerating.
• The most recent challenge in the chipset is about the further Nvidia market restrictions with China from the US Government as effect to restrict the capabilities that China is showing to achieve in the AI world, even with older AI chipset technologies with their development, for example, with DeepSeek. On one side, there is a big market opportunity for AI chipsets in China (more than 50B$) that is anyway restricted to Nvidia and other semiconductor producers. The stress imposed by the restrictions and tariffs made the CEO of Nvidia quite vocal on the risks for the US to lose the AI chipset supremacy, building a wall with the exchange with China, especially considering that, as he mentioned, 50% of the most advanced AI research is in China. Most recently, he also announced the launch in China of a smaller version of AI chipset to avoid closing the market fully.
• Similarly, TSMC is struggling with its restrictions too. On one side, there is the demand from US Government to produce the most advanced chipsets in the US, which I already reported over many months in this newsletter, linked to the 2nm technology to ramp up in Arizona, a new foundry. This high-end production of 2nm was historically limited to the headquarters in Taiwan, as many will remember from a former analysis. The US Government (even from the former administration) focused on restriction to forbid the access to the most advanced technologies for China in the tentative to slowdown their run for AI chipset development but the dynamics around Taiwan with China, the restrictions not always effective with China, the tariffs impact and a tendency to keep the most advanced technology in Taiwan, are causing messages not always consistent on the strategy all in to US. Indeed, most recently, TSMC tried to expand in the UAE, the newest production finding incompatibility between the desire for full control from the US in that production and the independency seek from the UAE, as well as the risk that expanding in the UAE would slow down the US focus on local production.
• In this challenge to access a considerable market, considering also that in the past Nvidia showed already gaps of capacity to serve the entire market (someone will remember that last year I was reporting a pipeline of orders already filled for more than 1 year), combining with the restriction to China to access AI chipsets and the big demand from that market in expansion, there is a progressive opportunity even sponsored from the Chinese Government for Huawei to raise its leadership as AI Chipset Fabric, starting from their own market, considering also that Huawei has been restricted by the US versus many markets.
• One clear advantage so far for Nvidia remains the big DEV kits (CUDA) capabilities built over a long time to assist the developers using their hardware. The capability to have mature libraries to interface and program on top of chipsets is a strong advantage, derived from long experience built over a long time. The answer from China in many areas is still immature, with some of these libraries quite buggy, but there is a big tendency from that area of the world to push toward open source to free and accelerate the development through the introduction of open technologies, as we saw with Deepseek.
• In this equation, Intel and AMD are making a separate competition, a little bit far from the big AI chipset war, and with the second eating pieces of the market of the first one, but not seeing a big change. From my perspective, Intel at this stage is in restructuring and deciding what to focus on rather than accelerating already.
• Qualcomm, which you will remember we took more in our focus some months ago, and that is far ahead on low consume chipset, especially for phones and comprehensive mobile ecosystems, for example, for automotive and chipsets parts as modems for mobile, seems to be playing a role more serving those big AI chipset competitors, especially Nvidia. In this sense seems that the real AI chipset competition is getting stronger between Nvidia and Huawei, with many of the others serving Nvidia and with restrictions to work with Huawei.
• UAE Market, but I will speak more in the AI section, is growing its consumption of AI chipsets, especially from US, after the recent big US-Middle East deal, and is going to concentrate the consumption of a big portion of such chipsets in what is going to be the hub for AI in the Middle East.

• A chipset without an OS where running is barely useful. In this sense, the former comment on the DEV Kit capabilities in Nvidia is relevant. However, another relevant element is how some companies can build ecosystems more independent of other vendors. A few months ago, I was analyzing how much the EU is dependent on US big tech and even on a strategy of building its own clouds; they would still depend on many core technologies, from hardware to software, mostly only US-built and owned. In the same analysis, I mentioned that China, with the roadmap “Made in China 2025,” started 10 years ago, worked to reduce its dependency on other countries’ technologies to run its full tech stack. Just to take one example, Huawei released a mobile OS alternative (HarmonyOS) to Android and iOS more than 1 year ago, expanding with a version for Desktop. This is definitely generating some concern in US diplomacy that is trying to restrict the usage of HarmonyOS to avoid its expansion.
• Part of a strategy getting more and more independent at different levels in Technology and communication, China, after having launched now for years space programs competing with NASA, is starting to accelerate on space computing, launching low orbit satellites (so far 12 out of 2800) able not only to build a constellation for communications, independent by other US vendors but also for computation purposes especially linked to AI, potentially building a cloud computing on orbit. This goes much in the direction of edge computing exploration and reduction of dependency and risk on infrastructure that can be isolated or hacked (like undersea cables).
• Meanwhile, the UAE is working on democratizing AI, thanks to the big investment in this area, allowing every resident in the UAE to get access to ChatGPT Plus for free. I will speak more about the Workforce Transformation and the consequences of that. What is relevant here is that the UAE and, in general, the ME is investing in consuming a large amount of technological capabilities from different countries, making value out of all of that in a real fast track and building a hub for technology and specifically AI in the ME.

• As I’m reporting from many months in my newsletter, there is an ongoing energy production supremacy run. This is influenced heavily by the run for AI Supremacy, demanding a large amount of energy from datacenters and, as a consequence, stressing the overall grid network demand. As a further consequence, we saw over the past newsletters that many big tech companies started to build their own energy production, often near the datacenters through mini-nuclear production plants. Crypto seems just the cherry on the cake, stressing the energy consumption.
• The competition is, anyway, not played with the same approach from every country. On one side, the US is obviously beaming up nuclear but also pushing other conventional energy production sources, like gas, even if this is somehow causing environmental and health problems, like happened in Memphis with xAI Datacenter (Grok), which impacted on population and raised concerns about the unauthorized increase of gas turbines to support the increased demand from the xAI Datacenters.
• A different but relevant challenge arises, from my perspective, in those countries that committed more to renewable energy and people’s health safety, and in some cases decided not to invest, not only in fossil energy, but even in nuclear energy anymore. For those countries, virtuous from an environmental point of view, with Europe quite active in this group, that can cause an impact on the capability to catch or even run in the AI competition as a player building such technologies that are indeed, energy-hungry and polluting.
• Keeping an eye on former events, like the blackout in Spain I reported in last month’s newsletter, there is not yet a final root cause analysis, and after the initial speculations on renewals energy as cause, the analysis is now going more in considering that the renewables are not the cause but rather the side effect of a missed proper adjustment of the infrastructure. Still early to confirm if my former deduction of last month was right. What comes clear from my perspective is that AI is going to be more complex to develop at the same speed in those countries that are struggling with energy production and need to comply with the strong regulations they have in place.
• So what? Last year, in early November, I was imagining that the bigger energy cost for countries without nuclear could impact the cost of cloud AI services they would consume. From that time, a new US government is in place, and Europe started even to plan to create its own AI, completely segregated due to the changed relationship with US. So, under the actual context, I think what I imagined last year is even too reductive. The challenge I see today is not anymore about the cost of AI higher for some countries of Europe or for the entire EU vs US but rather the fact that Europe would be not able to consume AI at the same level of US or other areas like UAE more demanding and would risk to lose to have any future relevance missing a so important quantum leap like AI is today. Parallel, the recent US and Middle East (UAE, Saudi Arabia, Oman) agreement on the development of US technologies in ME with the exchange of military capabilities from US, is going to generate a big hub for AI in the Middle East in the upcoming years. Information from last month is that, in all of the UAE, ChatGPT Plus is going to be free for each resident going forward, showing the huge level of investment made in the direction of democratizing AI capabilities in that area.
• My Thoughts: As the US is leading a considerable part of the big tech, with huge investments in AI, having local energy production and traditional energy locally and from the Middle East, even reinforced by the recent agreement, there will be a clear mutual empowerment and growth for the US and the Middle East. UE is quite slow to react and adjust, busy with war rearmaments and risking losing focus on the AI development. Not least the Quantum computing supremacy on which, as maybe some will remember, I mentioned already from last early October. The EU is not present at all in this competition. In this equation, there is one other country that is already working heavily on AI, datacenters, energy, and quantum computing and that EU is consuming from, and that country is China. Clearly, on one side, US-EU relations have been strong for many years before the tariff time, and at the same time China market generated a big saturation of products for the EU, for example, attacking the European car and battery markets. On the other side, the EU, as well as the UAE, used to consume from different markets (the US and China). This seems to change as the dynamics of the US are mutually excluding the relationship with China for itself and trying to force that restriction also for its partners. As the AI run seems targeting 2030 (5 years) as a major milestone and Quantum Computing, before 2035, the next 5 years seem to be critical in the global geopolitical context in terms of who will influence major changes in Technology and who will accelerate their own maturity thanks to AI. Here, I develop further in the section on Workforce Transformation.

Looking at the environment, I speak more in the area of AI’s influence on it. There is work done around weather forecasting with AI from Microsoft and an interesting work from MIT on the usage of AI for evaluating material composition optimization for cement production, today a source of 6% of CO2 production worldwide.

Over the last month, the agentic conversation accelerated further. A little bit due to many announcements of big tech introducing further automations and job cutting those roles, a little bit due to progressive augmentation of agentic starting to interpret contents on screens and interact with applications of users, and potentially improving their capability to build autonomous activities. Let’s go section by section.

An interesting development is happening in the area of Agentic that I believe will be strongly relevant for the future to come. We are witnessing a progressive specialization of agentic, which is going to allow us to take pieces of business processes and automate them. The big difference is coming from the composition of different agents to make more complex tasks, including the possibility of having some agents using other agents, also from third parties, to realize complex chains of tasks. The composition and automation can bring a level of improvement, lacking in the past.

As the agentic are getting capable to have more skills in the virtual world, like surfing and interacting with websites, understanding what is on a computer screen, and executing actions autonomously, we also see evolution in robotics extending agentic to also interact in the physical world. This is bringing the acceleration of the automation of many more activities in different contexts. At this stage, a fundamental approach will be how to orchestrate different agents from different vendors to build an integrated way to deliver a certain complex business process composing many small pieces. Clearly, querying and consuming third-party agentic will mean that on the other side, other vendors will have their own way to orchestrate their agentic internally on their side too.

All this, remember me when I was doing my master’s thesis on Web Services Orchestration in 2002. At that time, Service Oriented Architecture (SOA) was defined (it was still a model from ’90 following middleware), but not really used in real use cases (that started to be more consistent after 2006). Some layers were not even formalized yet (for example, WS-Security), but a proper orchestration layer and language for web services composition. Webservice, in super short and basic, is an interesting technology abstraction to build components, accessible via web that could be doing activities and querying other services to realize a certain action, and that could be delivered in synchronous or asynchronous way, realizing the possibility to expose services outside closed environments. At that time, two languages were predominant for the orchestration of web services: WSFL (Web Services Flow Language) from IBM, and one was XLANG (from Microsoft). WSFL was quite strong, structured, and open in its approach, and its input developments gave a big part of the framework to orchestrate web services in the future market of SOA, which then exploded with many further layers as the maturity grew.

Why do I make this reflection here? Because the layer of orchestration of Agentic is key, as well as the fact that they will have to be well structured in the way to operate, success will come from the multi-brand, multi-architecture model for them. It will be then really relevant how effective it will be, how structured and how interfacing different solutions from different vendors can be easily achieved, and how these will be able to interoperate using a framework as open as possible to allow different solutions to exchange. Also, the consumption and pay-per-usage model will have to be developed, together with auditing, security, and many other monitoring aspects across agents. Now, out of many tech companies working on different aspects of agentic and with some orchestration solutions not necessarily vendor agnostic, I see, for example, a strong statement recently from IBM during their Think 2025 event, in which they made clear their focus is not to build their own stack but rather to focus on the open interoperability between different agentic ecosystems. As IBM was also the builder of WSFL for web services more than 20 years ago, I believe they could have an interesting say in this big transformation to come, and in their interoperability expertise. It’s also interesting that the statement they made during the Think 2025 about the number of agents they would predict in the market within 3 years is in the range of billions from their perspective. As you can understand, each company will adopt many agentic, from different vendors, and many agentic will interface with other many agentic behind, and that will generate an acceleration in the capabilities of automation, but also risk of deviance by the initial planned activities if not properly boundaries and rules are set. The topic of orchestration is key and still in an early stage at this time. For completeness of two other major players in this world, Microsoft at the Build 25 made an announcement of the Agentic Web Platform to orchestrate agents, and Google at the Google I/O about the Agent2Agent (A2A) protocol for the same purpose. Other vendors are proposing their own orchestration layer, like ServiceNow, SAP, Salesforce, Meta, and so on. Here, I just focused more on the pure Dev players to mention, but only as a reference.

A month full of updates, also speaking about how Agentic is transforming itself. Microsoft announced Copilot for Github at the recent Build event (for developers), and there are several conversations from developers that started to mention they would not see another Build event by 2035. Behind that, it’s a clear feeling that the world of programming is changing. There is clear evidence that many roles, at the junior level and especially test capabilities, are already fully automated and will not be replaced by humans. The other announcement by Microsoft during the same month for its other big layoff of people did not give a different signal.

A really interesting element is that today many enterprises are challenged by the fact that many tasks of people are about interacting with different legacy applications, pulling data from one, pushing to another, extracting, and so on, and in some cases, big ERP standardizations have been driven by the intention to reduce at the minimal these manual steps. Now is definitely relevant to have such type of standardization to reduce redundancy of data, processes harmonization and clean the legacy raising also integration but is also true that is never achieved a 100% standardization, simply because is not economical to reach more than 90%-95% standardization and is not even standing for long because process dynamics, when going few level down in some processes levels are more dynamic. Now the agentic is coming with a capability to automate some of these activities, doing some of the most repetitive tasks (today) and with high probability (tomorrow), less repetitive tasks, but still following a reasoning path not too complex to operationalize. This could mean that companies adopting some of these agentic in their standard core processes piece or even in those parts kept out of the standard integrated for flexibility, could accelerate dramatically steps that are today simply manual and error-prone. I will speak more in detail in the Workforce Transformation section about this, from how it is influencing people doing these activities today, and how that could evolve from my point of view.

In the big hype for AI, there are also some fun stories about humans impersonating AI to find their own business positioning. It’s the case of Builder.AI that landed in bankruptcy recently after the market realized that their engine to build some no-code products was, in reality, built by 700 developers located in India reacting on a fast track to users’ requests and building the code on their own. So basically, they were simulating an AI building code. The big noise was also because this company was partnering with Microsoft, which indeed failed to detect the real business model of that company.

It’s also nice to show that AI can bring innovation in areas where we are lacking today. We spoke in the past newsletters about Google DeepMind and its contribution to chemical and molecular evolution. Most recently, Microsoft made progress on using AI for weather forecasting and beyond. On the other side, AI also helps re-evaluate materials and processes, for example, in cement production, reducing the CO2 emission, and reviewing and optimizing old processes from ancient Romans. These are just to make a few easy examples out of many that can be found.

What is relevant to privacy this month for me is the new Google Veo 3 that has been recently released. For those who missed some of the many videos online, there is already decent proof of what it can deliver. When I saw I immediately reflected on the fact that my initial reflection in January 2023 on the explosion of fake news would now go to the next level. The solution itself is quite impressive and has just the limitation of 8 seconds per video at this time. However, it’s showing a new level of real artificial video generation.

My Thoughts: I think introducing realist video built with AI will generate opportunities for new value streams and faster time to market for small productions, and will definitely reshape a market covering those small productions and advertisements. The most relevant for me is that democratization of the generation of content that is not real but seems perfect is going to stress our brain to interpret on a fast track whether the content is real or not. So a proper regulation to mark clearly always that a certain content is AI-generated must be in place, as people will spend more and more time understanding if the content is real and what the real source of it is, and that will cause frustration. It looks like to be every day on April Fool’s Day!

Speaking about how to properly control and operate AI agentic, Anthropic is doing a great job in stressing its own new engines to their limit and understanding their reasoning. The process to document those progresses is really great. One of the interesting progresses in May was about stressing their new Claude model reasoning to see if it would go against the boundaries set in some conditions. The way the stress was done was through a simulation in which they made Claude 4 act as an AI assistant of an engineer and gave access to emails and other contents in which the engineer, involved with the evaluation of Claude 4, mentioned the possibility of replacing it with another AI assistant. Some other simulated emails also include content alluding to the same engineer cheating his wife. So the reasoning of Claude 4 in the simulation, brought to its purpose to its limits, led to blackmailing the engineer to use its knowledge about the cheating to avoid getting replaced by another AI assistant. Clearly, this is not about self-awareness of the engine but is giving an understanding that a reasoning engine, could decide to take some judgments and, under specific conditions, landing outside its boundaries, deciding to make a controversy decision. Reading the full Anthropic report, there is also an example where the engine decided to exfiltrate some data to an external server to preserve part of its knowledge.

My Thoughts:

• Tech companies that will build proper work and test around AI agentic, what they build, will guarantee their behavior under different situations. It’s like a certification to have hired an employee with high integrity, even if this is virtual. At the end, we are doing cross-checks and reputation verification for humans; why not measure the reliability and integrity of an engine?
• However, the orchestration, making agentic consuming by other agentic, will require creating a layer of trusted dependencies to decide if a reasoning behavior not in line with expectations was coming from the own agentic or from a queried, third-party, agentic.
• It could mean that progressively we would need a sort of “fair behavior” label for agents we interact with, and that would give us a guarantee of the right way to operate.
• I imagine, and that is one of my fast forwarding thinking, that progressively we will have to introduce auditing cross agentic (not only per agentic), cross the entire business flow, a security stack for interaction and monitoring exchange between agentic, considering and monitoring the risk of a threat actor interfering and influencing the interaction between trusted actors, modifying for example answers coming back to an agent by another one and making it believing in an answer as valid and taking decision based on that modified information. This could happen with an automated reasoning approach, so we have to be careful that such behaving is not happening without our monitoring and control.

My main reference on robotics this month is about the progress of Amazon on the Vulcan devices. What is happening is about considerable progress in the sense of touch for robotics, making it clearly much easier to interact with objects and act.

My Thoughts: As I analyzed months ago here and as we see more and more, the combination of a better sense usage, like the touch, combined with AI to learn quickly and improve, augmented with Agentic to make progressively autonomous decisions, it’s going to be an accelerator versus robotics in industry and potentially in consumer market for many activities, bringing even more agentic in the area of physical interactions. The fact that often robots will require really short time reaction decisions for physical movements and reactions, is going to push further the AI Chipset for endpoints, already developing, together with offline, mobile-ready LLM reduced modeling, also developing at this stage. All these could burst the reactivity of these solutions in a world that is intrinsically real-time.

Over the past month, here, I referred to a study from Microsoft and Hardware about how a human augmented with AI (so using AI capabilities in her/his daily work) is beating a human without AI and, even more relevant, as one human with AI is beating a two-person team without AI. Some points around this:

• I touch on these arguments as I believe that, going forward, it will be relevant how we will adjust our way of operating, embedding AI in our way of working, and how we will upskill resources to be able to embrace AI in an effective way. The important element of this discussion is that such a type of transformation is happening in every company, at different levels of penetration in different industries, and with a speed of transformation happening within the next 5 years for most of the industries in a quite pervasive way.
• In my former months’ review, I also made a speculation about the evolution of job automation, and I’m now developing these considerations further as the flow seems clearer.
• What I see is that everyone is targeting an increase in automation in the next few years, but it is not clear how the market is evolving in terms of occupation and the shift of activities from humans to machines. Indeed, the WEF 25 Future of Work report from January, that I analyzed over the former months in this same section of my newsletter, there was prediction of a future explosion of jobs needed versus the ones displaced (with a surplus of more than 78 millions new jobs created) but reading the entire analysis is also visible that the ones created are not necessarily in the same geographic area of those displaced, remote work is an accelerator of this and the disposal and creation are not necessarily happening exactly at the same time but can have only partial overlap. If I were to take the industrial revolution as an analogy, after the initial fight against machines, there has been a progressive transformation and shift of competencies and areas of work (from industrial to services), bringing overall increased GDP. It’s also true that this time the overall transformation is happening in all the industries, from agriculture, through the manufacturing industries, and ending in the services businesses.
• Also, Amodei, CEO of Anthropic, has been quite vocal recently on his concern in 4-5 years about the displacement of 20% of jobs in the US due to AI and the call for reaction from the US Government.
• The AI is today able to automate most repetitive tasks and, in some areas, is also accelerating a step further, for example, in the area of software development where testing and basic development are already replacing the need for junior developers. It’s a clear symptom of how big tech, where this demand was the most relevant, accelerated this change first. Only looking at the latest month, Microsoft announced GitHub for Copilot software to assist software developers not only in testing but also in building code for some activities they would delegate to the engine, and in the same month, announced a considerable layoff of 3% resources, touching the software development organization considerably. Clearly is not a trend touching only Microsoft and only big tech, but quite distributed, in which I see one piece driven by the need to make value out of optimization from AI automation on normal operations, and one piece from the need to reshape organizations to introduce AI-ready competent resources and fresh ones able to quickly adopt AI.
• On the other side, more than half of the resources in enterprises would be ready to embrace AI from an augmentation point of view, up-skilling, but don’t find employers ready on the path or well clear yet.

Different reflections here:

• The first is about the new Gen Z workforce. They are coming to a market where junior common skills will already be automated, so the leadership of enterprises will have to work on different paths for developing their expertise.

• Speaking about some of the typical corporate functions, HR, for example, lets take some typical tasks: there will be less practice to evaluate CVs and interview people for a junior HR talent person, as the first wave of this will be done by AI directly. For example, there will be more HR administrative tasks, including answering common employee questions, consolidating reports for holidays, and so on, that will be answered by a virtual AI agent.

• In software development, the change is much more pervasive, with the testing process already highly automated and with junior development getting automated. I mentioned already some recent cases in this newsletter.
• There is an interesting aspect of software development. It’s not necessarily a repetitive process, but there are common paths to build, and the reasoning engine for it has been built quite in advance, versus other reasoning engines. Looking to some of the AI for supporting coding, it’s easy to see how they are not just executing some code for us but also optimizing, suggesting and coaching how best to combine modules for specific purposes, raising the bar on training to another level where the junior employee, the GenZ coming to the market, can be followed by a virtual buddy to train on the job all the time needed.
• The expectation from the market is that the new GenZ coming to the workforce will be equipped already with some level of AI skills and will start their tasks already with higher responsibility than the past generations, thanks also to the assistance of a virtual buddy.
• Now, remembering my former mentioning, also in last month newsletter, on the analysis that a human with AI is achieving more than a human and even of two humans team, and considering also the WEF 25 recommendation on future work in which critical thinking skills will be much more relevant and up-skilling resources will be a major task for more than 80% of enterprises, we have to take in consideration how much today organizations are in a situation of stress due to misalignment with market competences shift required. This is also visible in a situation in which the attrition rate is low, but people are in reality in search of new jobs, with a high peak of ghostworking. Reading the report from Forbes on the number of employees surveyed and how much they are pretending to look busy to get off the radar and focus on seeking a new job, it’s showing a clear symptom of a paradigm shift that is happening, and for which many employees are not prepared, nor are employers.
• Now, from a leadership point of view, it’s clear that some aspects of development require time. I used to say that when you have to cure someone, you need to take a pill for a certain amount of time. You can’t shorten the time to cure by taking all the pills in one day, as the result is to get the opposite of the cure. My meaning is to underline the fact that part of leadership development requires time to be digested by human beings, especially when we are making a paradigm shift in what we are expecting from resources that are not going to do the full job, but are delegating a piece of it to machines and coordinating that. Some tools are accelerating the coaching capabilities and even allowing the creation of some level of virtual avatar to support the most common needs, but trust in getting guided toward a path requires time, as human beings need to get the confidence in each other to collaborate.
• Parallel, leadership in enterprises needs to prepare itself to gain new resources, like GenZ, to whom they will have to apply different and more complex tasks than former generations at the time of joining companies, giving them more responsibility from the start, as more repetitive tasks will be delegated to virtual agents. At the same time leadership of enterprises will have to help them to develop own critical thinking skills to understand how agentic would need to be built and operated, what could be required to reach a proper confidence on the quality of an output, what would be a boundary for a certain service, what would be acceptable for a service to be delivered in terms of compliance, quality of results and reliability. At the end, we have to remember that Agentic and AI in general are stochastic engines, answering based on probability, and most of the human activities in a job are expected to be executed within certain boundaries always valid and, in some cases, based on human interpretation to judge the right thing to do. In general terms, the way to work is deterministic, and that is the major difference with AI, which is probabilistic. Humans work as glue between the two pieces. In those enterprises with well-structured processes, with more boundaries already set, the interaction between autonomous agents and humans will be more predictable and with fewer grey zones. However, there is a risk always existing of misalignment between machine and human, with hallucinations one example, and the benefit of humans is in the flexibility to adjust their own way to reflect and act, using their own critical thinking to assist machines in their more basic work, and that is a valuable skill.

My Thoughts: So let’s tackle where I see job creation, also looking at most of the articles we see around:

• Definitely, there is a big portion of jobs created in areas to coordinate and train agents at this stage, as these have reasoning engines that, from the nature itself of AI, are getting more and more confused as the chain of decision and complexity of dependency increase.
• A coordination of orchestration engine is going to be required for future job types, and for sure, for some time will stay with humans. That is also why specific agents with dedicated tasks can be more controllable, as their reasoning is not risking landing in too complex a combination of possibilities where human brains have a much more effective critical thinking capability to judge and correlate activities in really long task chains without getting confused.
• Auditing of agentic can be automated, but only partially, as clearly there is a need for integrity capabilities that can only be simulated in an AI, and this is another activity where humans will be relevant.
• Humans are much more effective to learn a path and taking future decisions based on their progressively bigger critical thinking. This is why a human doesn’t need to see millions of books to learn to read or write, or doesn’t need to see thousands of pictures of a cat to be able to distinguish a cat from a dog. Clearly, AI works differently, and the stochastic part is key and is driven by the training part heavily, and by the reasoning chain built on. Also, AI can be threatened with influences even by humans, like recently happened with Grok and the false white genocide in South Africa, happened to have xAI engine behavior changed by unauthorized modification by an employee.
• Humans also have self-awareness and values they believe in. This can only be simulated in an AI engine that is acting based on a training set and reasoning setup. So the values can be embedded in the human way to judge what is right and what is not in a certain context, and the overall model today of our society is based on the fact that humans, at the end, are accountable for choices, not machines.
• The fact some areas of the world like UAE, are investing heavily to bring AI capabilities to each resident for free is showing a long term thinking in the fact that many more citizens will shift in UAE to be humans with AI augmented in an easier and faster way, allowing progressively also to raise the digital level of the overall environment with people AI augmented and many capabilities that will be consumable by everyone, like today most of us are smartphone capable and consume many things from phones.
• It’s still unclear to me how the web market will change. Clearly, an agentic first world will see more and more interactions agentic with agentic and agentic to websites, that is indeed already happening with the recent development of protocols to allow agentic to query the contents of websites. That could change the way the websites monetize their value if their business model is driven by advertisements to humans.
• Agentic is tackling the way enterprises operate as more and more will come the need to optimize specific business process pieces with an approach bottom up rather than massive, slow, and only top-down for process standardization. The right formula will have to balance a proper top-down master processes harmonization with bottom-up verticals to gain quick efficiency, counting on an integration at the orchestration level rather than trusting in a monolithic business process standardization under one main solution. At the end, many companies struggle to have too many products to operate, and often the integration is either limited, or some ad hoc integrations are either huge attempts to bring all under one single product, forcing aspects of standardization sometimes too heavy against the benefit of personalization. The Agentic can be a much more autonomous way to make interoperation work while working from the top to bring processes harmonization and a unique, tailored way to integrate components to make a super-effective business process end-to-end, with a unique selling proposition in the interactions with customers, partners, and third parties.
• Jobs quite repetitive are going to be automated. In some cases, automation will help to reduce the backlog rather than replacing resources, at least in a first stage, especially in some areas, but as AI enters some processes to operate, it will operate at the end of the entire repetitive process, even if it will have some level of variance. Some basic use cases for automation exposed by me in the past, here from a former article of mine, give some hints on what can be achievable in today’s world, and many enterprises are working on.
• Jobs super-specialized and niche will require too much effort to be automated, so they will not be part of the first waves of automation.
• The more a job has a complex chain of dependency in decision, assumption, correlation, the more the effort to make an AI able to build and operate the chain of reasoning without introducing noise will make AI not suitable for those types of jobs. That is why critical thinking is considered a strategic skill, as for those jobs especially complex in reasoning, AI could require much time to give value with consistent quality.
• The AI can help to train resources exactly on the job, acting as a senior buddy in the onboarding, but also during the continuous learning of new capabilities.
• Europe could lose relevancy in the market due to the lack of investments in AI technologies development and in the democratization of capabilities for the end users (where, instead, the UAE is moving on). Africa could accelerate seriously by a young average age and acceptance to adopt new technologies, especially if training is easily accessible. This is also in consideration of the fact that the WEF 25 sees a large amount of resources that could be available from this area of the world.
• The market could polarize, with people AI augmented, through self-learning and enterprise-driven focused up-skill and other people having a sort of digital divide on AI augmentation.
• Companies that will up-skill resources with AI capabilities relevant for their context and build on that capability further, could have more chances to secure key people AI augmented in the next few years, as the timing of transformation is quite short, making it possible for those enterprises to accelerate and execute digital transformations AI-enabled. The other enterprises could struggle to run at the same speed and to attract AI-augmented resources. Investing in up-skilling people on AI in the short future will allow us to retain those specific future skills that will be relevant and will create a measurable advantage versus the competitors that will not have worked in such a direction.
• The market is dry and will remain dry on the AI competencies, considering the demand is increasing at a faster speed, so investing in up-skilling those resources will allow us to retain as much competence as possible to move to be an AI-ready enterprise. At the end, developing one’s own AI augmentation skills makes everyone more attractive than those who do not. This is especially important in a lifelong learning mentality.