The thesis
Artificial intelligence agents will replace the traditional software interface. In place of the screens, menus and input forms that every program today offers in its own way, a single interface will remain, the same for every task: a chat in which one says what one wants and, beside it, a space in which the resulting artefacts appear, whether a document, a table, a drawing or an order to approve.
We state it as a thesis and not as a prophecy, and so we treat it as a thesis should be treated: by declaring which observations it rests on, which trend follows from them, and what would have to happen for it to be considered wrong. Anyone looking for the announcement of a revolution can stop here; anyone who wishes to follow the reasoning will find the steps below, in the order in which we took them ourselves.
What the thesis claims, and what it does not
It is worth defining the terms, because a vague thesis can be neither defended nor refuted. By traditional interface we mean the set of screens that a program makes available to a person so that they can enter data and give commands, along paths that have to be learned program by program. The thesis claims that this layer is destined to lose its function, not that the software behind it is: the data, the rules and the calculations of an ERP, a CRM or a design tool remain as necessary as before, and if anything become more so.
What changes is the addressee. The program stops addressing a person and starts addressing an agent, which acts on that person’s behalf and returns a result to be read, corrected and approved. The human being keeps the request and the decision, the two things for which their presence cannot be replaced.
First signal: what we are asked for
The first signal does not come from a study but from our own work, and we report it for what it is: a direct observation, not a statistic. For some months the requests we receive have changed in nature. Ever more rarely are we asked for a new screen, and ever more often are we asked to make what an organisation already owns, namely its document bases and its business systems, available inside an agent.
The request takes two forms, which from the point of view of the thesis are equivalent. In the first, the agent is one of those running in AIDeskPro; in the second, it is the tool people have already installed on their own computer, such as Claude Cowork or ChatGPT Desktop. In both cases the person making the request does not want one more program to open, but the exact opposite: they want a body of information and functions to become reachable from the place where they already work, which is a conversation window.
Second signal: cost
For a single conversational interface to replace the others, automated reasoning must first of all cost little enough to be used for every operation, even the smallest. On this point a measurement exists. Epoch AI has calculated that the price needed to obtain a given level of performance from a language model falls on average by 47% every quarter, that is, roughly thirteen-fold a year, and that no other technology for which a historical series exists, from electronic computing to DNA sequencing, has ever moved at a comparable pace.

The figure should be read with two caveats. It concerns a fixed capability, so the most recent models remain expensive while what was the best a few months earlier becomes almost free; and the pace slows over time, down to less than five-fold a year for a capability that is two years old. Even assuming only this slower pace, the obstacle of cost has a predictable and short lifespan.
Third signal: the quality of the work
The second condition is that the work done by the agent be on a level with that done by a person, since an interface that produces results to be redone replaces nothing. Here too a measurement is available, GDPval, in which real tasks from forty-four occupations are given both to a professional and to a model, and an expert in the same field indicates which of the two deliverables they prefer without knowing who produced them. In 2024 the model won or tied one time in eight; in 2025 almost one time in two; in April 2026, in more than four cases out of five.

The limits of this measurement are part of the data and should be reported with it. The latest values are stated by a model maker and not by an independent observer; they concern single deliverables and not prolonged activities, in which models still succeed in fewer than half the cases; and models are not equivalent to one another, although field measurements show that, once a certain level is reached, the same result is obtained at costs that differ by tens of times.
Fourth signal: software opens the door
The third condition is that existing programs agree to be used by an agent rather than by a person. It is the most recent signal and, in our judgement, the most significant, because it depends on decisions taken by the vendors themselves. The protocol that makes it possible, MCP, was published in November 2024; the official library for adopting it went from fourteen thousand downloads in that month to 296 million in October 2026; and in the meantime Salesforce, Slack, HubSpot, ServiceNow, Google, Microsoft and Workday have made official access available to external agents, almost all of them in the course of 2026.

The list is not complete, and the absences should be declared as much as the presences: to our knowledge SAP has not done so, nor have the main Italian ERP vendors, around which it is for now third-party companies that build the connections.
The trend
Taken one by one, these signals admit other explanations. Taken together they describe a single trend, that of one interface for all tasks. The cost of reasoning falls until it is no longer a criterion of choice; the quality of the result reaches, for a growing share of office work, that of a professional; programs make themselves accessible to agents; and those who work ask, as a consequence, to be able to reach everything from the same point.
It should be noted that the four tendencies reinforce one another. Every program that opens up to agents makes the conversation window more useful, and every person who gets used to working from that window makes it less justifiable, for the next vendor, to keep only its own screens.
Our conclusion
We conclude that software, in the coming years, will have to be designed to be used by agents, and that the interface addressed to people will be reduced to two elements: the place where one asks and the place where one sees the result. This is the approach on which AIDeskPro is built: it presents the conversation on one side and the artefact on the other, and it makes an organisation’s document bases and tools available to agents, its own or those of others.
An honest thesis also states what would disprove it. We would consider it wrong if the fall in prices stopped before making the cost of an operation negligible; if quality on long, chained tasks remained at current levels; or if vendors, having opened access to agents, closed it again to defend their own interfaces. None of these three things, to date, is happening.
As for timing, we commit to a date: we believe that by mid-2029 the choice of a business management system will depend less on the look of its screens than on its ability to be used by an agent.
Sources: Epoch AI, “The plunging price of thought”, 22 September 2026; OpenAI, GDPval, September 2025 and April 2026; Agent Arena, Work category, data as of 13 November 2026; npm registry, package @modelcontextprotocol/sdk, read on 17 November 2026; vendor announcements for MCP servers.

