Artificial intelligence and the economy: why we need to prepare now
Reading time: 15 min
Artificial intelligence is no longer just a technological innovation, a writing assistant, or a new marketing tool. It is gradually becoming an economic infrastructure, capable of changing the way companies produce, recruit, organize work, and distribute value.
On July 13, 2026, more than 200 economists, artificial intelligence researchers, and technology leaders published a joint statement entitled "We Must Act Now ." Among them were 16 Nobel laureates, as well as figures from companies such as Google, OpenAI, and Anthropic.
The text is only a few sentences long. Yet its message is considerable: AI could cause, over the next ten years, an economic transformation of a scale comparable to or greater than that of the industrial revolution, but in a much shorter timeframe.
The authors do not claim to know the future precisely. Rather, they call on economists, policymakers, and technology leaders not to wait until the consequences are visible before starting to prepare for them.
A warning, but not a catastrophic prophecy
The "We Must Act Now" declaration does not predict the inevitable disappearance of work or the collapse of contemporary economies. On the contrary, it recognizes that artificial intelligence could generate significant productivity gains, foster scientific innovation, and improve living standards.
But it also highlights the risk of massive job displacement , increased concentration of wealth and a growing gap between the speed of technological development and that of the institutions responsible for regulating it.
The real issue, therefore, is not just what AI will be technically capable of achieving. It is determining how these capabilities will be used, by whom, and for what purpose.
The same technology can enhance workers' skills or gradually make them replaceable. It can disseminate knowledge throughout an organization or concentrate decision-making power in the hands of a few. It can create widespread prosperity or widen the gaps between businesses, professions, and regions.
The outcome will not depend solely on the performance of AI models. It will depend on investment choices, public policies, labor law, taxation, training, and corporate governance.
Does AI actually improve productivity?
The initial empirical work provides encouraging evidence.
In the study "Generative AI at Work," Erik Brynjolfsson, Danielle Li, and Lindsey Raymond observed over 5,000 agents working in a customer service department. The introduction of an AI-powered chatbot resulted in an average increase in productivity.
But the most interesting result lies elsewhere: the benefits were particularly significant for the least experienced employees . The tool allowed them to quickly benefit from formulations, problem-solving methods, and best practices generally acquired after several years of experience.
AI then acted less as a substitute than as a mechanism for transmitting the company's tacit knowledge. It captured certain practices from the best employees and made them accessible to others.
This observation opens up a positive perspective. When used properly, AI can reduce the time needed to develop skills, improve service quality, and help new employees reach expert levels more quickly.
However, it raises a question: what happens to human learning when beginners receive the answer directly without having to construct their own reasoning?
Immediate performance improvements do not necessarily guarantee the development of lasting skills. An organization may save a few minutes today while creating a technological dependency that will weaken it tomorrow.
The productivity paradox
The presence of powerful technology does not automatically translate into an immediate increase in overall productivity.
Erik Brynjolfsson, Daniel Rock, and Chad Syverson described this phenomenon in their work on the "modern productivity paradox." Despite the rapid progress of artificial intelligence, macroeconomic statistics do not always reflect the spectacular gains that have been predicted.
This delay is not necessarily proof that the technology does not work. It may correspond to a period of adaptation.
Major technological innovations generally require additional investments: redesign of processes, training of teams, evolution of information systems, reorganization of responsibilities and creation of new products or business models.
Simply installing a tool is not enough. The organization must be reinvented around its new possibilities.
The history of electricity provides a frequently cited comparison. The first electrified factories sometimes retained an organization designed for steam engines. They had replaced the energy source without revising the layout of the workshops or the flow of labor. The most significant gains emerged when companies redesigned their entire production system.
Artificial intelligence could follow a similar trajectory. Companies that simply add a chatbot to existing processes will likely see limited gains. Those that rethink information flow, validation steps, and task allocation can achieve more profound transformations.
Augmenting humans or seeking to replace them?
In "The Turing Trap", Erik Brynjolfsson warns of a particular direction in technological development: the desire to design machines that imitate human beings in order to replace them.
This ambition is understandable. Automation allows for cost reductions and the large-scale execution of repetitive tasks. However, it can lead companies to measure progress solely by the number of jobs eliminated.
Another strategy involves creating systems that complement human capabilities . AI can analyze more information, detect anomalies, suggest scenarios, or prepare an initial version. Humans then retain contextual understanding, judgment, responsibility, and the ability to make decisions.
This distinction is crucial in martech environments . AI can generate segments, suggest email subject lines, or identify atypical behaviors. However, it should not unilaterally decide on acceptable sales pressure, the ethical interpretation of sensitive data, or the appropriate response to a vulnerable customer.
QUOTE
Erik Brynjolfsson, “The Turing Trap”, Stanford Digital Economy Lab, 2022.
« When AI augments human capabilities, enabling people to accomplish things they couldn't do before, humans and machines become complementary. »
The debate is therefore not simply about the automation of a profession. It concerns the distribution of tasks within each profession.
Some activities will be automated. Others will be accelerated. New responsibilities will emerge around control, validation, data quality, explainability, and exception handling.
Why do the forecasts remain so different?
Economists are far from unanimous on the scale of the transformations to come.
Daron Acemoglu, in particular, adopts a cautious stance. In "The Simple Macroeconomics of AI," he argues that the macroeconomic effects could be significant but much more modest than some technological scenarios suggest. His central estimate results in limited growth in overall productivity over a decade.
His reasoning is based on a fundamental distinction between the technical ability to automate a task and its actual profitability.
A successful demonstration in a controlled environment does not mean that a system can be deployed on a large scale. Some tasks are easy to measure and standardize. Others require an understanding of the context, legal responsibility, a relationship of trust, or the ability to manage rare situations.
Economic gains also depend on the quality of the chosen uses. Automating a low-value operation does not necessarily create much value. Accelerating the production of mediocre content can even increase verification costs, informational confusion, and consumer fatigue.
Increasing production volume is not always synonymous with increased productivity. A company can send out more campaigns, produce more variations, and generate more reports without improving its financial results.
Real productivity means producing more value with comparable resources, not just producing more measurable items.
AI as a way to reduce the cost of prediction
Ajay Agrawal, Joshua Gans and Avi Goldfarb offer another interpretation in the book "Prediction Machines".
According to them, AI can be understood as a technology that significantly reduces the cost of prediction . It's not just about forecasting the weather, sales, or customer churn. More broadly, prediction involves using available data to estimate unknown information.
A recommendation system predicts what a user might enjoy. A scoring tool estimates the likelihood of a prospect making a purchase. An anti-fraud model assesses the risk of a transaction being illegitimate. Generative AI predicts the most relevant continuation of text, code, or an image.
When the cost of prediction decreases, companies can use it in more decisions. But this development simultaneously increases the value of other resources: data, judgment, goal setting, and accountability.
A prediction is not a decision. It does not specify what the company should do, what risks it accepts, or what consequences it considers legitimate.
The more accessible prediction becomes, the more crucial the quality of human and organizational judgment becomes.
Employment: Disappearance of jobs or restructuring of tasks?
Debates about employment often pit two simplistic scenarios against each other. In the first, artificial intelligence would massively eliminate jobs. In the second, it would be merely an additional tool, comparable to a spreadsheet or a search engine.
The reality will probably be more heterogeneous.
A job is made up of many tasks. Some are repetitive and easily automated. Others rely on negotiation, empathy, experience, responsibility, or knowledge of specific situations.
AI can therefore transform a job without making it disappear. It can also reduce the number of people needed to perform an activity, while increasing the demand for new skills.
Entry-level positions and roles within companies deserve special attention. When an organization automates the simplest tasks, it sometimes eliminates the very activities through which beginners learned the job.
A company that automates all the first levels of work must ask itself how it will train its future experts.
This issue directly concerns marketing, communication, IT development, consulting, and customer service. Research, initial writing, reporting, and preparation tasks can be largely assisted. However, they also constitute essential steps in professional learning.
Training can therefore no longer be considered a simple support for adoption. It must be integrated into the very design of the new processes.
" Wages are unlikely to increase when workers cannot claim their share of productivity gains. "
Daron Acemoglu and Simon Johnson, cited by the MIT Department of Economics, 2024.
The effects vary depending on the company.
Not all organizations will benefit from artificial intelligence in the same way.
Large companies generally have more data, computing power, legal resources, and specialized skills. They can train or adapt models, negotiate with suppliers, and more easily absorb the costs associated with experimentation.
Small businesses also benefit from AI services accessible in the cloud. However, they risk becoming dependent on a few platforms for their tools, data, and decision-making processes.
This concentration could strengthen the power of infrastructure providers and companies with the best-performing models. It raises questions about competition, sovereignty, data portability, and value sharing.
The authors of "The Macroeconomics of Artificial Intelligence," published by the International Monetary Fund, emphasize precisely this potential turning point. AI can lead to inclusive growth or to a further concentration of income and economic power. Institutions and public policies will help determine the path actually taken.
What companies should prepare now
Preparing does not mean predicting exactly what models will be capable of doing in five or ten years. It means building organizations capable of adapting to different scenarios.
The first priority is to map tasks rather than just jobs. The company must identify which operations can be assisted, which can be automated, and which require human validation.
The second step must then measure the gains beyond the time saved. Quality, error rate, customer satisfaction, team learning, resilience, and control costs must also be taken into account.
The third priority concerns the retention of skills. When a system takes over a task, the organization must ensure that it retains sufficient internal knowledge to detect an error, regain control, or change suppliers.
Finally, governance must precede incidents . Rules relating to data, validations, intellectual property, traceability and liability cannot be improvised after deployment.
In a martech environment, these principles notably involve knowing:
- what data can be transmitted to a model;
- what content needs to be validated by a human;
- how are the segments and scores controlled;
- who remains responsible for an automated decision;
- how to measure errors, biases and false positives;
- what procedure to apply when a system produces an unexpected result.
The goal is not to systematically slow down innovation, but to make its adoption sustainable.
Prepare the institutions as much as the technologies
The "We Must Act Now" declaration is also addressed to governments and institutions.
Educational systems will need to train students not only in the use of AI , but also in critical thinking, verification, and understanding the limitations of models. Labor law will need to address new forms of monitoring, evaluation, and the allocation of responsibilities. Taxation will need to adapt if a growing share of value is captured by technological capital and platforms.
Competition policies will also need to address the concentration of infrastructure, business models, and data. As for social protection systems, they will need to be able to support potentially faster and more frequent career transitions.
These measures do not assume that the most extreme scenario will occur. They are a form of risk management.
When a transformation may be profound but its scope remains uncertain, inaction is not a neutral position. It amounts to letting market decisions and existing power dynamics determine the trajectory alone.
Do not choose between optimism and pessimism
The work of Brynjolfsson, Acemoglu, Korinek, Agrawal and many other economists does not converge towards a single forecast.
Brynjolfsson highlights AI's ability to augment the workforce and disseminate best practices. Acemoglu cautions that macroeconomic gains should not be overestimated and that not all forms of automation create the same value. Korinek examines the potential consequences for labor, inequality, competition, and governance. Agrawal analyzes AI as a reduction in the cost of prediction, which is gradually transforming the entire spectrum of economic decision-making.
These approaches are less contradictory than they appear.
They show that AI has significant potential, but that this potential will not spontaneously translate into collective progress. Between the capabilities of a model and a real improvement in living standards lie businesses, educational systems, institutions, rules, and political choices.
In conclusion: a transformation to be guided rather than endured
The main lesson of the "We Must Act Now" declaration may not be that artificial intelligence will necessarily eliminate millions of jobs. It is that the decisions made before the transformation is fully visible will have a decisive influence on its consequences.
Expecting complete certainty would be illusory. When statistics unambiguously confirm a breakdown in the labor market, it may be too late to quickly rebuild the necessary skills, training systems, and redistribution mechanisms.
Artificial intelligence is not an external force that will mechanically impose a single future. Its impact will depend on how organizations choose to automate, share productivity gains, and preserve human capabilities.
The question is therefore no longer simply: "What will AI be able to do?"
It becomes: "What type of economy do we want to build with it?"
Some references
- “We Must Act Now: A Statement on AI’s Transformation of the Economy” – Stanford Digital Economy Lab and signatories – Stanford Digital Economy Lab – July 2026.
- "Generative AI at Work" – Erik Brynjolfsson, Danielle Li and Lindsey R. Raymond – The Quarterly Journal of Economics – May 2025.
- “Artificial Intelligence and the Modern Productivity Paradox: A Clash of Expectations and Statistics” – Erik Brynjolfsson, Daniel Rock and Chad Syverson – National Bureau of Economic Research – November 2017.
- “The Simple Macroeconomics of AI” – Daron Acemoglu – National Bureau of Economic Research – May 2024.
- “The Turing Trap: The Promise & Peril of Human-Like Artificial Intelligence” – Erik Brynjolfsson – Dædalus, American Academy of Arts and Sciences – spring 2022.
- “Prediction Machines: The Simple Economics of Artificial Intelligence” – Ajay Agrawal, Joshua Gans and Avi Goldfarb – Harvard Business Review Press – April 2018.
- “The Macroeconomics of Artificial Intelligence” – Erik Brynjolfsson and Gabriel Unger – Finance & Development, International Monetary Fund – December 2023.
- “A Research Agenda for the Economics of Transformative AI” – Erik Brynjolfsson, Anton Korinek and Ajay Agrawal – National Bureau of Economic Research – September 2025.
