A famous line from Samuele Bersani’s song Giudizi universali uses a description — “you are just a copy of a thousand summaries” — that, until recently, could also have been used quite effectively to describe language models and chatbots such as ChatGPT. Their use was mainly based on a very simple dynamic: a person asks a question, the system processes the information available to it and provides an answer. A model that was certainly promising in terms of personal productivity, but still far removed from the concepts of delegation and automation that characterize work within an organization.
The summer of 2026 made a shift that was already underway much more evident: chatbots are becoming technologies to which tasks can be delegated. On 9 July, OpenAI launched ChatGPT Work, describing it as an agent designed for longer and more complex activities, capable of gathering and analysing information, working with connected applications and files, and ultimately producing documents, spreadsheets, presentations, reports and other finished outputs. The distinction introduced by OpenAI itself is significant: Chat remains the most suitable environment for obtaining a quick answer, developing an idea or working on a specific part of a problem; Work, by contrast, is designed to complete tasks from start to finish.
Almost at the same time, OpenAI transformed its application directory into a plugin directory. A plugin can include operating instructions, specific capabilities and, above all, applications through which ChatGPT can access external services and their related data and actions, within the limits of the permissions granted. Google Drive, Slack, Gmail, Canva, the Adobe suite and other software can therefore become part of the context in which the AI operates.
It is difficult not to see an analogy with the path followed by Anthropic with Claude and its connectors. The underlying logic is similar: a language model becomes more valuable when it is no longer isolated within its own conversation window and can access, subject to the appropriate permissions and the security assessments carried out by company IT departments, the tools in which the business actually stores and updates its information. Claude, for example, can connect to services such as Google Drive, Gmail, Notion and Canva and use this information as part of the work it has been asked to perform.
In any case, the common direction is more important than the differences between individual platforms: chatbots are no longer simply standalone applications into which users upload data and write prompts. They are becoming an operational layer positioned above software, shared folders and business processes.
For a company seeking to expand into international markets, this transformation is particularly relevant. Export activities constantly require the ability to bring together information from different sources: market trends, local competitors, trade fairs, distributors, regulations, enquiries received by email, commercial data stored in the CRM, technical documents, price lists, presentations, sales reports and competitors’ activities. The problem is rarely a complete lack of information; more often, it is the fragmentation of that information.
This is where the difference between Chat and Work becomes particularly clear, a distinction that has recently also been adopted by Claude and Copilot. With a chat, we might ask: “Who are the main German competitors of our company?” We would probably receive a useful initial piece of research. With Work, however, the assignment can become much more sophisticated: “Analyse our website and the commercial materials stored in the ‘Export Germany’ folder. Identify the five most relevant competitors in the German market. Analyse their positioning, product ranges, differentiating factors, listed distributors and the latest developments published on their websites. Compare this information with our catalogue and prepare a document for the sales management team highlighting opportunities, risks and three actions that should be explored further.”
This is not simply a matter of writing a longer prompt. The nature of the request itself changes. In the first case, we are asking a question; in the second, we are assigning a task.
Research, collection, comparison, analysis, organization and production of the final deliverable become part of the same workflow. Work can proceed through several stages, request clarification where necessary and ask the user to approve actions that require authorization. In this sense, the most important development in generative AI does not necessarily concern the quality of the answer to a single prompt, but rather its ability to maintain context and continue working towards a defined result.
A further transformation is taking place at the same time: prompts no longer necessarily have to start when a person writes them. ChatGPT’s scheduled tasks — a feature also available in other leading chatbots — make it possible to programme recurring activities and monitoring instructions. The features available depend on the plan, connected applications and permissions, but the underlying principle is particularly interesting: a prompt can become a process that is executed repeatedly over time. Scheduled tasks are now also available to eligible Free and Go accounts, while tasks triggered by specific events within connected applications require plans and configurations that support Work.
For an export manager, this means moving from “search today for what my competitors are doing” to “every Monday, check whether the specified competitors have launched new products, updated their catalogues, announced partnerships, participated in trade fairs or entered new markets; if you identify significant changes, prepare a summary and flag only what deserves attention.”
This is by no means a minor shift. In the first case, the initiative remains entirely human: someone has to remember to perform the research, open ChatGPT, formulate the request, check the results and distribute them. In the second case, intelligence activity becomes continuous. The value no longer lies only in the AI’s ability to search for information, but in the possibility of transforming a recurring information need into a routine.
Consider international competitor monitoring. An Italian company selling industrial machinery in France could create a weekly task to analyse a predefined group of competitors. The agent could check for new products, pages dedicated to specific sectors, participation in trade fairs, new certifications, content published on websites and other accessible sources, as well as significant changes in positioning. The result would not necessarily need to be a long report. It could instead be an executive summary highlighting only what has changed since the previous review.
The same logic can be applied to distributors. A company could periodically ask the system to monitor signals relating to potential partners in a specific country: new brands represented, expansion of product ranges, participation in events, organizational changes, acquisitions or the opening of new offices.
The same applies to market monitoring. Rather than carrying out an in-depth analysis once a year, companies can maintain a light but continuous form of monitoring, asking the AI to report only events that exceed predefined relevance thresholds.
An important principle emerges in this context: a good agent does not necessarily need to produce more information. It needs to reduce noise. “Notify me only if something happens that could affect one of our commercial decisions” may become a more useful instruction than “send me ten news stories about Germany every day.”
The next major step occurs when AI can use not only the Web but also a company’s internal information systems. This is where plugins take on a strategic role. Connecting ChatGPT to a CRM, a document repository, business intelligence tools or shared files enables the agent to link external and internal information.
Imagine a company exporting to five European countries. Its CRM contains data on commercial opportunities, shared files contain market analyses, distributor requests arrive by email, while orders and margins are stored in the ERP system. Individually, each source provides useful information. Greater value is created when those sources can be connected.
An assignment could therefore be: “Analyse commercial performance over the past six months in the French and German markets. Compare it with the pipeline available in the CRM and with the latest reports from the sales managers stored in the Export folder. Identify any discrepancies between sales performance, opportunities in the pipeline and the qualitative assessments provided by the sales team. Highlight the three anomalies that deserve further investigation and prepare the questions to be discussed at the next export meeting.”
This type of activity is difficult to reduce to the traditional idea of a chatbot. It is closer to a form of information coordination. And precisely for this reason, it raises an issue that companies should address before even choosing a platform: what data can the AI read? What data can it modify? Which operations can it carry out autonomously? Which ones require approval?
We can distinguish three levels. Activities such as research, collection, comparison, classification, initial analysis, organization and preparation of materials can be delegated. Interpretation, strategic evaluation, priority setting and the production of content in which context and tone are particularly important should be supervised. Finally, activities that have significant external consequences should remain subject to approval: sending certain communications, modifying information in company systems, making commitments towards customers or distributors, publishing content and taking commercial decisions.
Another particularly interesting area for international companies is what we might call document monitoring. Companies operating internationally constantly manage documents that change over time: catalogues, price lists, technical data sheets, certifications, sales presentations, contracts, marketplace guidelines, distributor materials and regulatory documentation.
An agent could be given the recurring task of checking a shared folder dedicated to a particular market and identifying which documents have changed. It could compare different versions, highlight inconsistencies and produce a list of required actions.
Suppose, for example, that a company changes a technical feature of a product in its main catalogue. That change may also need to be reflected in the French product sheet, the German distributor presentation, the Spanish price list and the materials uploaded to a marketplace. Traditionally, someone needs to remember where that information has been used. A document agent, however, can help identify the affected materials and flag inconsistencies.
A task could be formulated as follows: “Every Friday, check the commercial documents related to product line X. If a technical specification, price, certification or product feature differs between the master catalogue and the materials intended for the French, German and Spanish markets, create a report of the inconsistencies. Do not modify the documents automatically. Indicate the file, the section to be checked and the version you consider to be the most up to date, explaining the reason.”
The principle is important: the agent researches and prepares; the person decides.
This also suggests a possible evolution in the organization of export activities. Rather than having a single general-purpose agent, we could imagine several specialized agents: one dedicated to competitive intelligence, one to distributor monitoring, one to commercial documentation, one to checking opportunities in the CRM and one to monitoring regulatory sources for a specific market.
They should not be imagined as virtual colleagues with complete autonomy. It is more useful to think of them as intelligent processes to which companies assign a scope, sources, frequency, relevance criteria and escalation rules.
This distinction is what makes AI adoption particularly interesting from an organizational perspective. For a long time, attention focused on creating the perfect prompt. People were taught to provide context, define the objective, assign a role to the AI and specify the format of the output. These remain useful skills, but they are now only one part of the problem.
When a prompt runs every Monday, uses data from company systems, compares documents, produces a report and can potentially trigger a subsequent action, it is no longer simply a prompt. It becomes a stage in a process.
This is probably the most interesting meaning behind the concept of “loops”: rather than an isolated interaction in which a question and an answer complete the task, we have a sequence in which the AI receives an objective, gathers information, carries out activities, verifies the result and continues until the assigned task has been completed.
For companies seeking to operate in international markets, the implications are significant. AI can reduce the cost of acquiring knowledge about a market, increase the frequency of analyses that were previously carried out only occasionally, help a small export team manage a larger number of countries, reduce the fragmentation of information between sales, marketing and management, and make company data more accessible even when it is currently underused because querying it requires technical expertise or excessive amounts of time.
However, simply installing ChatGPT Work is not enough to make all this happen. An increase in individual productivity does not automatically translate into greater organizational productivity. An employee may use a chatbot effectively and save half an hour when preparing a report, but if they have to wait two days to receive a file from another department, if information is scattered across folders, emails and systems that do not communicate with one another, if documents have no clearly defined owners or if no one has established which version is the official one, the overall benefit remains limited.
AI does not automatically solve organizational bottlenecks. In some cases, it simply makes them more visible.
Before building an export agent, companies therefore need to ask questions that have little to do with artificial intelligence itself: where does the information originate? What is the official source? How often should it be checked? Which events are significant enough to trigger an alert?
The starting point should not be “What can ChatGPT Work do?”, but rather “Which repetitive, information-intensive and document-related activities are currently slowing down our ability to understand and manage a market?”
Starting from the tool risks producing demonstrations that may be technologically impressive but organizationally irrelevant. Starting from the process makes it possible to identify where AI can remove friction: a report that currently takes two hours every week, a manual check across twenty websites, information that exists within the company but is never used in meetings, or documents that need to be reviewed periodically.
The real leap forward is not moving from a simple prompt to a more sophisticated one. It is moving from “asking AI for something” to designing a process in which people, agents, data and applications work together according to clearly defined rules.
The next competitive advantage will therefore not come solely from the ability to use artificial intelligence better than competitors, but from the ability to organize data, responsibilities and workflows in a way that allows artificial intelligence to work effectively.