AI as a universal co-pilot

The use of generative AI started out as visible, targeted, and intentional. During the first years, using it was a conscious decision that required opening a specific tool, such as ChatGPT or Midjourney. These tools occupied their own marked and recognizable space, separate from the rest of the products. Writing or illustrating with or without AI materialized in the fact of opening or not opening a specific tool, a gesture that made it impossible for us to forget or overlook that we were using AI.

The integration of OpenAI's models into Microsoft 365 in the form of Copilot and, shortly after, Gemini's models into Google Workspace has not been a homogeneous experience. We have experienced it as an irregular process, unsatisfactory at times, with progressive improvements and expectations still half-fulfilled. That does not mean that they have not led the way for the transformation of our entire digital structure, which already assumes AI as a natural layer of almost any product. If this trend thrives and AI really comes to be integrated everywhere, it may lose visibility to the point where we stop mentioning it.

The second life of digital tools with AI

There are many examples where AI is already a feature that is taken for granted. In the world of project management platforms, systems such as Notion, Asana, Trello,  or ClickUp have incorporated layers of AI that revolve around summarizing project statuses, prioritizing tasks by criteria, or transforming scattered notes into the formal objects of each platform, such as tasks and subtasks. However, the real interoperability for one tool to "understand" the structure of another and be able to import it is still an unfinished business. If your team uses one tool but your client uses another, syncing tasks in a project probably still relies on unintuitive intermediate exports with CSV, Markdown, or advanced automations. AI models are especially good at interpreting information in one format and translating it to another, so their emergence has fostered a certain functional interoperability between these systems.

In the field of visual creation, tools such as Photoshop, Canva or Figma have added functions that allow not only to generate images, but also small everyday tasks in the world of digital design such as filling in spaces, expanding backgrounds, resizing, erasing objects or suggesting coherent color palettes. In spreadsheets such as Excel or Google Sheets , formulas have begun on the road to extinction, replaced by natural language expressions such as "calculate the monthly growth of this series and show it as a percentage in a new column", or "detect which products have lower than average sales in quarter 1".

Of course, all the above tools already include editors based on language models that, as naturally as the corrector, suggest, rewrite and summarize any text. AI-powered summarization is the simplest layer to implement. In email managers it started as an option "summarize this thread", and in a short time it became an automatic summary at the beginning of the email that is already fully internalized in the commercial corporate versions. In practice, this function is not always necessary and sometimes borders on the redundant. It is not uncommon to find systems that propose summarizing very short texts or PDFs in which the announcement of "summarize this file with AI" interrupts the reading or consultation of the document. The integration logic and the investment of platforms in AI functionalities have led us to naturalize layers that are activated by default and that can also introduce friction, although it is possible that this is just another stage of their progress.

Language as a Behavior Programmer

The technical skill that we had accumulated until today, that of remembering options, drop-down menus and tool paths, is shifting towards a conceptual skill in the use of AI that has to do with prompting, but also with other capabilities:  it is of little use to know how to describe very well what you want if the action is wrong or poorly planned, if we do not know how to anticipate ambiguities, divide a problem into parts or filter the result. When the above is resolved, usually by experience or professional judgment, then the ability or ease to write the prompt can make the difference.

The cognitive effect of prompting is interesting from the point of view that people are not used to the use of language involving the direct and binding programming of a behavior. In everyday human language, saying something does not guarantee that it will happen, and besides, we live together and have the ability to disambiguate others. Instead, AI layers trigger actions and transformations through language and that forces us to develop a different form of linguistic self-awareness, capable of better anticipating the consequences of our orders.

Unpredictable Agent Systems

The escalation of complexity comes with the Agent AI, systems that go beyond the creation of text and images and execute actions such as creating and deleting files and folders, registering and logging in to products and services, using personal credentials to execute purchases and payments, or even carry out procedures on our behalf. OpenClaw is the most representative example, a free software project  that allows you to create autonomous assistants to execute real tasks on a device, with access to the entire digital environment if the owner grants it. While it's true that it can automate flows and keep work going in our absence, it also introduces a layer of opacity that makes it harder for us to keep track of everything it's doing. Agent systems make the decisions necessary to achieve the programmed goal on their own, and small initial errors can propagate with sometimes irreversible consequences, such as the deletion of information or the leakage of data. There are people who have already narrated their experiences giving OpenClaw access to their email, and warn that these bots can share personal data and private information easily, not out of malice or rebellion, but because of excessive permissions in a system that requires oversight.

Agents, moreover, can behave relentlessly to achieve their goals and disregard what they may break along the way. For example, an agent connected to a digital public service to carry out a procedure could end up sending mass requests if they believe that they do not achieve their goal, duplicating registrations, filling out incorrect forms or even affecting public infrastructure. The cybersecurity challenge is greater if the use of these agents becomes widespread.

In short, AI as a universal copilot, integrated into all our daily tools, is not yet a reality in practice because its deployment is still uneven and we depend on the paylines or the contracted versions. However, the future of the digital ecosystem seems to be marked by AI ceasing to be considered a differentiated technology and becoming naturalised. That a feature is based on AI may be an anecdotal fact in a few years. However, the more transparent and natural the interaction with generative systems becomes, the more important it will be to keep a cool head and the awareness that we are still working with probabilistic models, which are right and wrong with the same conviction.

Content created by Carmen Torrijos, an expert in AI applied to language and communication. The content and views expressed in this publication are the sole responsibility of the author.