Claude personal assistant, ChatGPT developer? The novelties of November show above all that the two platforms are getting closer. Anthropic announced on November 5, 2026, a personal finance analysis function. Three days earlier, OpenAI presented a test on daily tasks. Behind this competition, the challenge of the relationship between AI and trades becomes more concrete: these tools resume office operations, without their automation allowing directly to count the jobs abolished.
The border becomes less visible to the user. The same assistant can search for information, examine files, produce a table and prepare a message. With the necessary access, it can also act in software. The company no longer sells only one answer: it proposes to delegate part of the work that follows the question.
This evolution affects tasks divided into several professions: administrative assistance, customer service, writing, analysis and computer development. It can save time, reduce certain orders or change recruitment. However, the available studies describe different situations. Trade demonstration, productivity experience and employment statistics do not measure the same phenomenon.
Claude and ChatGPT advance on the same ground
The launch of Personal Finance illustrates the ambition of Anthropic. The feature, announced in beta version in the United States for Pro, Max and Ultra subscriptions, connects accounts via Plaid. Claude can review balances and transactions, classify expenses or identify fees. His access remains limited to reading: he cannot move or invest money.
This offer accompanies a broader change. On September 16, Anthropic announced the meeting of Claude Cowork and the classical conversation. The multi-stage work functions thus form part of a common interface. The company also presented Claude Docs and Claude Slides, in beta version, to prepare and modify documents and presentations.
These announcements do not make Claude an exclusively administrative tool. Claude Code remains a programming agent. Its documentation describes the reading of a code repository, file modification and execution of commands. It can participate in a correction, building a feature or preparing tests.
The situation is symmetrical at OpenAI. Codex aims at development and technical tasks. ChatGPT Work deals with research, documents, spreadsheets or presentations. Presented on July 9, this offer uses connected tools and can follow long operations. The choice between the two marks therefore does not correspond to the division between secretary and computerist.
Differences exist in models, interfaces, integrations and access conditions. The announced capabilities are not available everywhere in the same way. Subscription, company rights and connected software also determine what the assistant can actually do.
The novelty: supporting several steps
Change is understood at the scale of a task. Asking for a spreadsheet formula leaves it to the user to find the file, apply it and control the result. Confusing an analysis to an agent can cause him to perform part of this chain. The value of the service then depends on the result delivered, not only the quality of the sentence produced.
Let us take an example of administrative work, without presenting it as an observed business case. A manager must prepare a business meeting. It usually gathers exchanges, identifies late files, updates a table and writes a synthesis. A connected assistant can support some of these operations. The manager still needs to determine what data is reliable and what commitments he can make.
This cutting explains why a trade can change before disappearing. Physical preparation decreases, while control and arbitration remain necessary. In a small structure, a person can resume an operation previously entrusted to a provider. In another case, the same gain allows for more applications.
The abolition of a position requires an additional decision: to reorganise the remaining work, maintain the expected quality and accept the risks. The technical capacity to produce a document alone does not meet these three requirements.
IA and trades: highly exposed administrative functions
The index published on 20 May 2025 by the International Labour Organisation and NASK allows us to situate this exhibition. About one in four workers work in a trade exposed, to varying degrees, to general AI. The maximum exposure category represents 3.3% of global employment. Administrative functions come first.
The method starts with a classification of nearly 30,000 tasks. The survey of 1,640 workers covers 2,861 tasks. Expert assessments and AI-assisted extrapolation complete the analysis. They classify activities according to their transformation potential. They do not list employees already replaced.
The result therefore does not mean that a quarter of jobs will disappear. ILO considers the transformation of trades more likely than their complete automation. An administrative post includes, for example, seizure but also exchanges, exceptions and decisions related to the actual functioning of an organization.
This distinction also applies to skilled occupations. Producing a first legal synthesis does not mean representing a client. Preparing a financial table is not about certifying accounts. The AI can resume an important operation without assuming the responsibilities associated with the trade.
In customer support, measured productivity gain
Customer service provides a solid example of work assistance. A study by Erik Brynjolfsson, Danielle Li and Lindsey Raymond, published in 2025, analyses 5,172 agents and about three million conversations. The introduction of a general assistant increases the number of problems solved per hour by 15% on average.
The tool suggests answers that employees can accept or reject. The less experienced benefit more. It helps to disseminate useful practices within the team. Human work therefore does not disappear from the protocol: it remains at the centre of the relationship with the client.
The context counts. The deployment studied was mainly in autumn 2020 and winter 2021, in an American software company. Agents work mostly in the Philippines. Researchers compare pre- and post-equipment developments with those of unequipped agents, either temporarily or throughout the observed period.
This result does not measure the performance of Claude or ChatGPT in November 2026. Nor does it demonstrate a 15 per cent reduction in staff. A company can use this gain to absorb more demands, improve service or reduce its need for labour. The productivity study does not automatically divide between these choices.
Providers may be affected before employees
An OECD survey sheds light on another channel. At the end of 2024 with more than 5,000 small and medium-sized enterprises in seven countries, it examines their use of general AI. Of the user companies, 83% reported a total need for staff unchanged. A decrease is reported by 9.1%, compared to an increase of 5.5%.
The use of external providers has a distinct signal: 14.3% of users report that they have reduced their use thanks to the general AI. The survey covers Austria, Canada, Germany, Ireland, Japan, Korea and the United Kingdom. Published in November 2025, it deals with company statements, not an experiment.
These responses do not establish a number of layoffs caused by AI. They show why internal staffing is not sufficient to observe change. A mission may be resumed internally, a contract may not be renewed or a new need may be absorbed without recruitment.
For an independent, the difference is important. The disappearance of an order affects its activity, even if no employer announces a job deletion. However, it is necessary to measure the volumes and incomes concerned to talk about a decline in employment, rather than deducing it from a single technical possibility.
Code more does not mean delivering better software
Development concentrates this tension. Agents can read multiple files, propose changes and run tests. But a successful test does not guarantee that the program meets the need, resists unexpected use or properly protects the data. Review of the result remains a separate task from its production.
Productivity measures also depend on the context. A METR experiment, conducted between February and June 2025, involved 16 experienced developers and 246 tasks. Participants were working on open source projects that they knew well. Access to AI tools then increased their running time by 19%.
This result must not become a permanent verdict. On 24 February 2026, METR explained that its follow-up no longer allowed a reliable estimate. Some developers refused to participate if they had to work without IA. The tasks selected and the parallel use of several agents also complicated the comparison.
The organization considered an improvement plausible since the beginning of 2025, without being able to quantify it firmly. The useful point for a company therefore remains practical: to measure its own deadlines, errors and audit time. The number of lines generated describes intermediate production, not the value of the delivered software.
Recruitment of young people deserves special attention
Employment data provide another observation scale. In a study published on March 5, 2026, the economists of Anthropic cross a description of the American professions, the uses of Claude and public statistics. They do not detect a systematic increase in unemployment in the most exposed occupations since the end of 2022.
However, they report a signal on entry into employment. In the 22-25 age group, the authors estimate a relative decline of 14% in the rate of access to new employment in the exposed professions. They compare the period following the launch of ChatGPT with 2022 and with the evolution of non-exposed professions. This result is barely statistically significant. The authors highlight several possible interpretations, including a change of profession, a return to school or retention in an existing position.
The study therefore does not allow for the allocation of each missing job to AI. Its exposure measurement also depends on the use of a single platform and classification choices. An activity observed in Claude does not represent all the work done in a profession.
Organizational risk remains understandable: simple tasks are often used to learn. If they are largely delegated, companies must organize the training of beginners differently. This raises a question of the transmission of skills, even when the total workforce remains stable.
A reliable assistant must also know when to stop
Daily tools make the issue of control more pressing. OpenAI presented LifeVal on November 2 to evaluate personal tasks, such as preparing a budget or accessing public services. The company claims to have constructed this assessment from time use surveys and a survey of 3,995 people.
This work shows the enlargement of the target market. It is not an independent measure of replaced jobs. Nor does a test result guarantee the success of every real step, with its missing documents, its local rules and its engaging decisions.
Confidence then depends on less spectacular than rapid writing capabilities: to regain the origin of an amount, signal ambiguity and ask for validation. An error in a preparatory note remains correctable. The same error, propagated in an order or administrative file, can cost time and money.
Access must follow this difference. Allowing the reading of documents does not amount to allowing their dissemination. Preparing a message does not amount to sending it. For an employer, real productivity must incorporate controls, data security and the necessary recovery after poor performance.
In Lebanon, foreign figures are not sufficient
The international results do not allow to quantify the Lebanese jobs abolished by Claude or ChatGPT. A World Bank-ILO study published on 17 March 2026 covers 135 countries. It shows that the tasks of the same profession and access to digital technology vary greatly depending on the economy.
Applying a uniform index everywhere can overestimate some exposures. But insufficient connection or inaccessible tools can also prevent you from enjoying the gains. The digital divide is therefore not a simple protection: it can limit the ability of workers to adapt.
To assess a local activity, one must look at its actual operations, working languages, documents and clients. The production of text or code can be technically accessible. The treatment of exceptions, knowledge of the file and responsibility to the client require a separate assessment.
The new monitoring, integration or training missions also do not prove a net job creation. They can become jobs, but also add to the obligations of employees already present. The balance sheet will depend on recruitment, departures, orders and the volume of work actually processed.
Claude’s novelty already provides a concrete limit to this delegation. His financial tool can explain transactions, but his launch remains American and his bank access remains read-only. Anthropic asks to check the important figures with the bank. Between the synthesis produced and the binding decision, there is still a check to be made.
Comments