Quo nos ducis, IA?   Where are you taking us, AI?

By Carlos Cosials

It would have been more popular to ask “Quo vadis, AI?” (Where are you going, AI?), but that would have distorted my deep conviction regarding the true nature of AI. I am often bothered by how others—inevitably, when “addressing” it via prompts—attribute anthropomorphic qualities to it, asking it to respond like a human and, worse, reacting as if it surpasses us when it does so.

So, I will not make the act of questioning it the central focus, nor will I dismiss it (even though I maintain my stance of viewing it as an “artifact”). However, I will grant it enough “substance” to acknowledge that it possesses affective capabilities—meaning it exerts influence or affects us, rather than possessing emotions in the human sense.

In short, AI is affecting us, and we cannot gauge the extent of that impact because we do not know where we want it to take us. It will be a new era—different, with both ups and downs—for there is no technology without ideology, as there is always an underlying purpose.

Thus, it falls to us—much as we consultants habitually do—to: (1) determine where we stand, (2) identify where we are headed (or where we will arrive), and (3) map out what that transition will look like.

We approach this with the humble intention of sharing questions rather than certainties, set against the backdrop of the summer of 2026—a time when news on the subject has been relentless, offering a dizzying array of weekly changes and updates. Consequently, by the time you read this, some of the information may already be obsolete.

Where do we stand?

We have completed the bulk of automation, a process that entailed the application of digital technologies. In other words, the introduction and adoption of ICTs allowed everything capable of being digitized—shifting its substrate from “natural”(1) to artificial—to transform into this new nature, thereby altering how we interact with and manipulate information in the pursuit of efficiency. The digital realm enables us to go further and faster at a negligible cost compared to physical equivalents. We have fared well and made progress; the downsides have been outweighed by the gains, making the overall outcome beneficial. While some sectors or activities have vanished, looking back now, we can affirm that, on the whole, it has been a positive development.

Companies have had to adapt under the banner of what has been termed “Digital Transformation”—though, with a broader historical perspective, one could argue that businesses are constantly adapting, and those that fail to do so simply perish. Thus, the climate is ripe for technological adoption, just as it has always been. In November 2022, a new company—OpenAI—emerged on the market, offering consumers “ChatGPT” and creating a new product category. This occurred even though the underlying technology was already known in academic and some business circles, and was already present in Google Translate, which had adopted it in 2019 (2). It created the same magical impression that Google Search had produced years earlier.

What was the essential difference? The mechanistic determinism that technology companies had previously offered in their results and outputs. This was the case despite Google initially presenting itself with the slogan “Always in beta” to highlight the constant novelty it provided—though that label was also sometimes interpreted as a sign of product immaturity.

In any event, a search result was reproducible in retrospect, and—*a priori*—the algorithm generating the result was consistent given the same content and context. In other words, the Google Search product manager could be held accountable, as they were able to explain why the product delivered specific answers rather than others. That is precisely why the original Search product manager resigned following the introduction of AI into the search technology; they could no longer take responsibility for the results it produced.

Thus, by comparison, we find that the new AI capability is non-deterministic—a characteristic stemming from, and made possible by, its stochastic foundation.

In other words, all the advantages that statistics offers as a technique apply here, yet with the inherent limitations of statistical modeling. This means it is not a technique that fully models behaviors, but rather one that provides an approximation based on its sample data.

Thus, what we currently know as AI is an attempt to deliver a result based on statistics—which is precisely why it feels plausible—yet without the ability to trace back the reasoning used to construct it or reconstruct the provenance of the information provided. In short, we are dealing more with beliefs than with established facts; this calls into question a significant part of the scientific method that has brought us this far—the principle that a demonstration supporting a thesis had to be reproducible and, therefore, explainable.

But what is AI?

Image: iStock by Getty Images

In short, it is a two-word acronym that has overshadowed everything that came before it; in the media landscape, it has successfully captured the general public’s (B2C) attention, creating the illusion of magical results thanks to its anthropomorphic nature.

It appears to respond and reason like a human sitting behind the screen. Google Search used to respond with incredible speed, finding what we were supposedly looking for—its main virtue being that it spared us the tedious task of searching. It pointed us to key references but did not—until recently—construct a synthesized, coherent answer for us. It is no longer merely an ordered list of references; instead, it “responds” to the intent behind the query (or “prompt”) we submit.

It is important to highlight that, with AI assistants or chatbots, there is a crucial step preceding the response: the interpretation of the query (or prompt) written by the user. In other words, it shifted from Google Search’s obsession with simply processing the phrase provided by a “lazy” user to discerning the underlying intent behind those words. Take, for example, the query “headache upon waking.” Interpretation here is tricky: is the user looking to buy painkillers online (transactional intent), or are they seeking an explanatory medical article (informational intent)? It is worth noting that, over time and through hard-learned lessons, search engines have come to prioritize medical safety in such cases.

That said, given the current landscape (as of September 2026), AI offerings can be broken down into four categories. While these categories have distinct features, they do overlap in terms of capabilities and adoption history:

1.         Inference Engine
2.        Conversational Assistant
3.        Operational Agent (Delegate)
4.        Process Orchestrator

Inference Engine: refers to Machine Learning—specifically the construction of stochastic inference models designed to “anticipate” future behaviors based on the analysis and understanding of similar past situations. Examples include recommendation systems and anomaly detectors.

Conversational Assistant: refers to the chatbots that emerged in late 2022. We can query them on any topic—without needing to provide prior input ourselves—and expect a response that is constructed, well-argued, and reasoned. Operational agent: refers to a superior capacity—beyond mere solicitation—to issue commands when our “desire” seeks the execution of an action, shifting from awaiting a response to securing the result of the requested action: “A restaurant reservation,” for instance. In my view, the term “agent” is an Anglicism that translates awkwardly into our Latin context; until recently, we have always perceived an “agent” as a representative of an authority to whom the exercise of that power has been delegated. However, as individuals, we have delegated—rather than “agencied”—our authority to other people; consequently, delegating it to artifacts is even harder for us to grasp.

Image: iStock by Getty images

Process orchestrator: refers to the establishment of a set of agents that interoperate to achieve a goal greater than what any single constituent agent could accomplish on its own. This process also involves setting necessary, precise safeguards and constraints—or, to use the technical term currently gaining traction, “guardrails.” As of today, this represents the highest level of sophistication attainable through the integration of all the aforementioned characteristics.

What will AI be?

Given all the above, a key question arises: will it be an infrastructure or an application/platform?

This is analogous to what we witnessed with the Internet itself—which was co-opted for their own business interests by companies already operating in the telecommunications sector—and its “over-the-top” applications, such as Google, which did not invested in transmission infrastructure, but rather in storage and computing—leveraging these to create a new business category at a higher level of abstraction, establishing a “layer above,” much like a search engine.

With this model in mind, players in the conversational and agentic AI arena are seeking to scale their offerings—moving from individual or B2C solutions to enterprise or B2B models. This shift allows them to create value beyond their conversational inference “engines” (or models, as they are known) and thereby expand their capacity to drive demand for computing power.

In this context, I would like to revisit the automotive analogy: the “lever” needed to move the vehicle is the existing engine (exploitation), but when it comes to exploration—determining the destination—we need to set a course. That is the role of the steering wheel, the mechanism used to guide the vehicle. Ultimately, AI will take us to a destination shaped by the combination of power and direction.

That said, there is a key aspect of this technology’s application that will have a fundamental impact beyond the workplace: the academic sphere and how we will learn in the immediate future.

It is undeniable that the way we learn—or, at the very least, how acquired knowledge is assessed and certified—is changing. Society needs to understand the level of competence possessed by the certified individuals presented to us.

There are no clear indications yet as to how this issue will be addressed, though all signs point to a shift in emphasis: prioritizing the development of a student’s independent judgment over the mere transmission of knowledge. Education is fundamental here, as it lays the groundwork for decision-making and responsibility—qualities that are subsequently applied within organizations and society at large. In short, AI will serve as a new (inference) infrastructure upon which specialized applications will be built—applications capable of extracting value worth paying for.

As we are seeing with both “hyperscale” operators (OpenAI, Anthropic, Google, SpaceX, etc.) and startups, a range of applications is emerging that focus on solving specific problems; in these cases, mastery of the context enables the delivery of a valuable proposition. For instance, Anthropic—with its Claude product and its “Code” specialization—is creating a specific application geared toward generating code in programming languages. Meanwhile, DeepL has specialized in translation, leveraging AI (including generative AI) to provide its service.

And finally, what will the transition look like?

It promises to be painful, as it will affect the elites—defined here as those societal groups granted authority through their mastery of knowledge. Their status stemmed from the fact that the capacity to reason and argue was once their exclusive domain, a result of the effort they had invested using their own intelligence. That exclusivity has now been surpassed—not yet universally, perhaps, but the scope is expected to widen significantly due to the massive adoption of ICT.

In other words, the ubiquitous deployment of digital technology—acting as a superior entity that “governs” the course of events (granting loans, driving vehicles, etc.)—means we are dealing with an entity (computer systems) that responds to “commands” received in the form of programmatic code written in programming languages. If that code can be generated by a generative inference engine—responding to a prompt that may itself have been generated by (another) agent—we face a scenario that is, to say the least, unsettling. If we add to this the lack of retrospective traceability and the stochastic nature of this monstrosity, we have a “weapon”—one with shaky foundations but possessing a recursive power that is impossible to gauge beforehand. This explains the “pause” (or the call to slow down) now being demanded by the very creators of this monstrosity—though the true reasons may well be economic (erecting barriers to entry) rather than rooted in moral or ethical concerns.

Image: iStock by Getty Images

Furthermore, the impact of its adoption will be felt across organizations in general and businesses in particular. Many of these are built upon a pyramidal structure of accountability—that is, the capacity to respond to events in a responsible and traceable manner. Yet, we are now seeing that we can ask AI to summarize a situation and—almost—provide a diagnosis ready for action. So, why not delegate the action itself? Indeed, this is already happening with the granting of small, quick loans, for instance.

In any case, the reality is that no one wants to be left behind; the application—or rather, the adoption—of this new AI is being piloted, with a profusion of experiments, tests, and pilots everywhere—offering a glimpse of the true impact and allowing one to gauge the actual gains against the costs and repercussions involved in their adoption..

^[1] We would consider a vinyl record or a CD-ROM “natural” in the sense that it is a tangible object—and therefore one we can “manipulate” with our hands and touch.

^[2] https://research.google/blog/recent-advances-in-google-translate/