The Impact of Personalized AI on Brazil’s GDP

Executive Summary

• AI personalization is a cross-cutting economic driver: Personalization based on Artificial Intelligence (AI) affects practically all productive sectors – not just technology.
• Significant macroeconomic impact: The simulation of a scenario in which AI personalization would be prohibited indicates a 1.64% drop in Brazilian GDP by 2035, equivalent to R$855 billion in accumulated loss in 10 years.
• Most affected sectors: Industries that have a greater relationship with other sectors would suffer the greatest impacts — such as oil refining, metallurgy, and agriculture

—, since losses spread throughout production chains.

	THE 10 MOST IMPACTED SECTORS	dVA
1st	Manufacture of tobacco products	-1,913
2nd	Oil refining and coke plants	-1,897
3rd	Slaughter and meat products, including dairy and fish products	-1,886
4th	Biofuel manufacturing	-1,883
5th	Sugar manufacturing and refining	-1,873
6th	Agriculture, including agricultural and post-harvest support	-1,871
7th	Other food products	-1,856
8th	Livestock, including support for livestock farming	-1,849
Gº	Oil and gas extraction, including support activities	-1,821
10th	Electricity, natural gas and other utilities	-1,813

• Why this matters: The study demonstrates that AI personalization is more than a convenience feature, but also critical infrastructure for productivity. The data reveals that severe restrictions on the use of this technology would not only penalize the digital sector, but would generate systemic losses of competitiveness in basic industries and Brazilian agribusiness.

Introduction

What is Personalization?

Defining the object of research is a complex task: in information science, we call this concern ontology – the effort to build systems to classify and represent knowledge. In the case of personalization, it is a particularly challenging task, as it involves technical, economic, communicational and even cognitive dimensions.

A systematic review published in 2022 analyzed 383 academic studies on personalization and concluded that there is no clear consensus on what personalization is (Chandra et al., 2022), as the term can be interpreted from different perspectives: as organizational strategy, technological capacity, human-system interaction, or user experience – just to name a few.

In this study, we preferred to adopt the communication process perspective, close to that proposed by Blom (2000): personalization is the process that changes the functionality, interface, content or distinctiveness of a system to increase its personal relevance for an individual.

Although personalization is mainly associated with digital systems, its existence predates the digital age: has been present in commercial and communication practices since the end of the 19th century, starting with marketing direct mail, including credit bureaus and credit card loyalty programs, among other strategies.

When a merchant separates products that a repeat customer might like and decides to present them to them when they arrive at the store, this action is also a form of personalization.

Personalization is not an isolated phenomenon of the digital economy, but a structural trend throughout the history of marketing: a reflection of the effort to convert human preferences into data and, at the same time, optimize attention, efficiency and engagement.

Chandra et al. (2022) also investigated different ways of classifying types of personalization, noting that Most studies organize based on six criteria: (i) what data is used; (ii) what is personalized; (iii) how it is customized,

(iv) who customizes; (v) where the data comes from and (vi) how it is communicated to the customer. In the context of this study and the Brazilian regulatory reality, we can explore these categories as follows:

CUSTOMIZATION – CLASSIFICATION FORMS

WHAT DATA IS USED

Non-Personal Data: in this case, we can talk about the personalization of systems for business use, in which data unrelated to natural persons is used to personalize the system (the most common model in industrial and B2B applications);

Personal Data: when the system uses data related to a natural, identified or identifiable person.

WHAT IS CUSTOMIZED

Content: adjustment of information, media or materials displayed to the user, aiming for greater informational relevance and engagement. Examples: social media feed, email recommendationsstreaming, playlists on music services, recommended articles on a news portal.

Advertising and Propaganda: targeting advertisements, commercial offers or political messages. Examples: targeted digital ads, newsletters, exclusive promotions in apps, political campaigns, financial service offers.

Services and Products: dynamic adaptation of interfaces, language and functionalities of products or services, tailored to the profile and behavior of each user. Examples: applications that rearrange menus, games that adapt the gaming experience, chatbots and virtual assistants that adjust the tone of language, software with dashboards customized, wearables that offer health alerts.

Price and Access Conditions: modulation of the conditions of access to goods, services or public policies, such as prices, fees or benefits. Examples: differentiated credit rates, autuarial analysis.

HOW IT IS CUSTOMIZED

(Kwon and Kim, 2012)

Mass communication(one to all): the system makes the same offer to everyone, based on general average preferences. In this case, there is no real customization, but only targeting based on general usage statistics.

Group segmentation (one to n): The company divides the public into groups with similar characteristics (also known as clusters). This is, for example, the most common model of digital advertising – offers are rarely for you, but for groups of similar characteristics.

Individual customization (one to one): in this case, personalization is done for each specific person, based on data, history and behavior – like a recommended list of movies on streaming.

WHO CUSTOMIZES

System Provider: adaptation is driven entirely by the organization that controls the system. In this model, business rules unilaterally define the user experience. Examples

include assignment of credit limits by financial institutions or layout of shelves in supermarkets.

User:customization results from a manual and deliberate action by the individual himself, determining to

priori how the system should behave or present information. This includes choosing options when purchasing vehicles, or configuring dashboards in ERP systems.

By the Provider, through the User’s agency: intermediate space, in which technical execution is performed by the system and calibrated by explicit signals of user agency (such as feedback signals, control adjustments, or corrections of output). Digital advertising

and like/dislike tools on email services

streaming fall into this category.

WHERE DOES THE DATA COME FROM

First-party data: Information collected directly by the organization during user interaction with its own channels and

contact points. Example: service streaming which recommends movies based on viewing history and user ratings on the platform.

Third-Party Data:Information obtained from sources outside the direct relationship with the user, such as data brokers, commercial partners or public sources, used to enrich profiles and infer behaviors. Example: A banking institution that uses credit scores provided by bureaus external.

HOW IS IT COMMUNICATED

In this case, a rigid classification would confuse rather than help, since, in practice, what there is is a transparency spectrum: one continuum which extends from

a pole from total opacity to the normative ideal of total transparency, with reality lying in the multiple intermediate gradations of this spectrum.

• at the extreme of opacity, the “zero-transparency” is currently incompatible with market demands – which demand information openness to the user – and with personal data protection legislation, which expressly prohibits it.
• at the same time, the absolute transparency It appears more as an axiological horizon than as a factual reality, since the technical complexity of certain systems and the cognitive overload imposed on the user also make perfect intelligibility unattainable.

The interaction between these six dimensions produces distinct effects, and risks are not an intrinsic property of an isolated category, but the result of how they combine. A practice based on intensive personal data is not always the most sensitive, and the use of aggregated data is not always harmless, the difference is less in the type of technology and more in how these dimensions meet.To illustrate, let’s compare two scenarios:

• a streaming system can utilize a user’s granular behavioral history (personalization one-to-one with first-party date) to recommend a movie. The ‘risk’ here, in case of failure, is a frustrating entertainment experience;
• On the other hand, a credit model might not use any individual data, but set the interest rate based only on your zip code or age group (targeting by clusters). Although it seems less invasive, the material impact is infinitely greater;

In practice, starting this study by explaining all this matters because the term “personalization” has become an umbrella used to justify everything from legitimate innovations to questionable practices. Without distinguishing what, who and how is personalized, what data is used and where it came from and how it is communicated, the public and regulatory debate becomes shallow., decisions end up treating very different situations as if they were the same and we lose the opportunity to see where there can be real gains in efficiency and relevance for the user – a topic that gains special importance in the context of the popularization of artificial intelligence and machine learning techniques.

AI and Personalization

In today’s world, Artificial Intelligence (AI) takes on the role of facilitating large-scale personalization, allowing you to adjust products, services and experiences. Very common examples include video or music recommendations, search engines that adjust results for each person, chatbots that adapt the language to the user’s profile, and page versions that change depending on browsing history.

We have countless examples of studies on the topic from this year alone: Nisar (2025) analyzed how AI transforms the interaction between consumers, companies and services, focusing on recommendations and engagement. The Future of Privacy Forum (2025) report, “Concepts in AI Governance: Personality vs. Personalization”, distinguishes how the

“Personalization” is addressed in the main Generative AI systems. Shafik (2025) addresses the impact of AI on digital marketing: machine learning and natural language processing techniques allow individualized personalization, overcoming clustering, while Bitra (2025) explores the regulatory challenges resulting from this scenario.

Even though the debate seems recent, different sectors have already adopted mechanisms for adapting content, products and services to user preferences for decades. What changes with Artificial Intelligence is the scale and depthof this process: techniques machine learning makes possible to customize not only what is offered, but also the behavior of the systems themselves that interact with the user. Thus, personalization ceases to be a peripheral function and becomes the central operating logic of various digital products and business models.

For the purposes of this study, we understand AI technologies as software systems capable of processing and analyzing data through algorithms and mathematical models, employing statistical and computational techniques to identify patterns, make predictions or generate new content. In other studies published by Reglab, we distinguish Analytical AI systems, aimed at solving delimited problems, such as fraud detection or demand forecasting, and Generative AI systems, capable of creating texts, images or codes from large volumes of data. Unless there is a distinction expressed in the text, this research understands both systems together, within the comprehensive concept of AI technologies.

Economic and Regulatory Relevance

Time is money. If a tool (a hammer, an excavator or software) helps you do a task faster, you save time, and We call this gain productivity growth. And AI studies have consistently shown how these technologies increase productivity.

Brynjolfsson, Li and Raymond (2025), Dell’Acqua et al. (2023), and Noy and Zhang (2023) were some who documented significant gains when people start using Generative AI tools in their daily activities. Eloundou et al. (2024) estimated that about 80% of the US workforce could have at least 10% of their tasks impacted by the introduction of these technologies – 19% could have more than half of their tasks impacted, and about 15% of all tasks in the US could be completed significantly faster while maintaining the level of quality.

These numbers suggest that AI is not a niche technology that will only affect a few specialized sectors, but rather a general-purpose technology that will have broad impacts across the economy.

A part of these gains comes from the ability of these tools to adapt to each user – i.e. your personalization. When AI knows your way of working, it delivers more useful and faster results. Without this, you lose part of the advantages it offers.

But the lack of this customization can cause problems that go beyond working slower. Recent research, such as that conducted by Gadhvi et al. (2025), found that the Non-personalization can lead to a 17% reduction in worker satisfaction and a 10% reduction in focus on the tasks performed. These numbers may seem abstract at first glance, but they have concrete, measurable implications for productivity.

EXPLAINING IN PRACTICEImagine that the AI ​​learns how you like to write emails – whether you prefer a more direct or formal tone, whether you usually start with a specific greeting, whether you generally end with “kind regards” or “hugs”. Over time, the suggestions she offers come “in your style”. Two colleagues asking the same thing will receive different answers, each tailored to their individual needs. This saves time – and increases productivity – because you don’t have to rewrite as much.

When we lose focus and switch from one task to another, part of our attention becomes stuck on what we were doing before – and this harms our performance because our mind cannot completely change focus on the fly (Leroy, 2009).

With non-personalized AI tools, this happens all the time: you need to reinterpret responses, make adjustments, and complete what the AI delivered. If the tool knows how it works, many of these adjustments would not be necessary.

And the problem goes beyond losing focus. Mark, Eudith and Klocke (2008) show that jumping between tasks increases stress, frustration and tiredness. In other words: non-personalized tools not only make work slower, but also more exhausting, less satisfying and, consequently, less productive (Oswald, Proto and Sgroi, 2015).

Finally, there is the issue of time. Dillon et al. (2025) found that using AI for emails can reduce the time spent on this task by 31%. This would free up almost four hours a week — time that can be used on more important activities. In a 40-hour week, this represents a productivity gain of almost 10%.

Potential Effects of Non-Personalization of a System

Attention residue

When part of your mind gets “stuck” on the previous task even after changing activities. It hinders your focus on the new task

10% loss of focus (Gadhvi et al, 2025)

Fatigue and satisfaction

Jumping between tasks and correcting errors is more tiring and frustrating. The more worn out, the less you produce. When you are satisfied, you work better and more.

17% loss of satisfaction (Gadhvi et al, 2025); happier workers produce up to 12% more (Oswald, Proto and Sgroi, 2015)

Time and productivity

Time saved on repetitive tasks can be used on important work. Less wasted time = more productivity

Reduction of time spent on emails by up to 31% (Dillon et al, 2025)

Given this body of evidence, Would it be possible to estimate what the effect of personalizing AI systems would be in economic terms? To answer this question, we first need to establish a methodology that allows us to translate these individual productivity gains into aggregate economic impact.

Methodology

SUMMARY – How This Study Was Done Comparison Scenario: The study simulates an extreme case, as if there were a law that did not allow AI systems to use data to adapt to the user. Reference value: To measure this impact, we use a comparison number: with personalized AI, the work yields 10% more; without customization, yields 10% less. Furthermore, we would have a 10% loss of focus. Together, these two effects mean a net loss of about 1% in total productivity.Economic model: The economy works like a network in which everything is connected. If one sector loses efficiency, it affects the others. Therefore, the study uses a general equilibrium model, which shows how the loss at one point spreads throughout the economy. Difference between sectors: Not everyone uses AI in the same way. The study crossed data on professions and tasks with IBGE statistics to calculate which sectors would be most affected by the ban on personalization — and how much this would weigh on Brazil’s GDP.

Methodologically, this study has four relevant aspects: a counterfactual scenario, a calibration parameter, the econometric model and the AI use exposure calculation.

Counterfactual Scenario: Non-Personalization

To estimate the macroeconomic effects of personalization, the model starts from a limiting hypothesis: a complete ban on personalization. This is not a prediction that such a scenario will materialize politically, but rather a methodological need to establish the “total exposure” of the Brazilian economy. This choice, although apparently extreme, is methodologically necessary.

EXPLAINING IN PRACTICEImagine a ruler: to measure any intermediate degree, we first need to define where the ruler begins and where it ends. This is why general equilibrium models require clear quantitative anchors. Trying to directly model hybrid scenarios – such as opt-in (prior consent) or opt-out regimes – would require making behavioral assumptions about how many users would accept or reject personalization and other elements that in themselves would represent a new study. By setting the scenario at the maximum restriction (total ban), we calculated the “ceiling” of the economic impact. This value therefore serves as a base reference: any regulation or judicial/administrative decision that partially restricts personalization (whether due to consent frictions, sectoral prohibitions or technical limitations) will have an economic cost that will be a fraction of this total impact.

So, the extreme scenario works as a stress test: it reveals the full size of the slice of value that is at stake. If looser regulation reduces the efficiency of personalization by just 20% or 50%, policymakers can use this study to proportionally infer costs, knowing that the real impact will lie on the gradient between the base scenario (full use) and this counterfactual scenario (no use).

Initial impact size: 1%

To estimate the economic impact of personalization carried out by AI systems, we start from reference values adopted as a calculation basis – something common in economic studies that simulate scenarios that are not yet observable.

In this case, and based on the studies summarized in Table 1, we consider that customization generates an average productivity gain of 10%, resulting in time savings and greater precision in tasks. This also means that the absence of personalization would cause a 10% productivity loss, associated with dispersed attention, lower user satisfaction and less efficiency in performing tasks.

In practice, it is like comparing two production lines: one equipped with tools adjusted for each worker and another that uses the same standard for everyone – assuming that the first performs its tasks 10% faster than the second.

Combining these gain and loss effects, the study simulates a –1% net productivity shockon tasks that could benefit from AI personalization.

EXPLAINING IN PRACTICEImagine that the initial productivity is 100. The use of personalized AI generates +10% → productivity goes to 110. If customization is prohibited, there is a loss of 10% on this new value (110) → 10% of 110 = 11. Final result: 110 – 11 = GG.

That is, productivity drops from 100 to GG, a 1% net reduction. It is worth remembering that this value is not directly observed in reality, but it is our reference to estimate the aggregate impact on the economy as a whole. And even so, it is a conservative estimate.

• This is because the calculation assumes that the negative effects of non-personalization only partially reduce the gains from AI, when in practice they can be greater, and

until it gets in the way of carrying out a task. Imagine an autocomplete system that, with each sentence, suggests incorrect terms that you need to delete

and rewriting – that is, the tool not only stops helping, but starts getting in the way, consuming correction time.

• Furthermore, the model does not incorporate other benefits of customization, such as learning specific contexts in Brazil, regional differences, or reducing “rework” in complex activities. A generic AI may write a flawless contract but fail to ignore the specific tax information of the contracting companies.

Therefore, the 1% shock should be interpreted as a minimum limit of the potential impact – a prudent way of measuring a phenomenon that is still developing. It is important to highlight that, given the differences between sectors of the economy, this initial shock could also have different values between them: that is why we assume that the degree

of exposure to the use of AI also takes into account these differences and, therefore, the resulting shock in each sector is (i) less than 1% and (ii) particular to each sector, as presented later in the text.

The Quantitative Model: General Equilibrium

Now that we have the scenario defined and the mathematical parameter, we need to translate this into macroeconomic impacts – that is, how much this represents, in reais, for different sectors of the Brazilian economy – which are not isolated from each other, but interconnected by a complex network of contractual and economic relationships. To capture these interdependencies, the most appropriate model is the one that we call it the general equilibrium model.

EXPLAINING IN PRACTICEImagine the economy as a large network of companies that buy and sell from each other. An automobile factory, for example, doesn’t just produce cars – it also buys steel from steel mills, electronic components from specialized manufacturers, tires from rubber companies, and so on. If something affects the productivity of the automobile industry, it not only impacts the cars that come off the assembly lines, but also all the companies that supply inputs to this industry. Likewise, if productivity in steel mills falls, this makes steel more expensive, which in turn affects not only the automobile industry, but also civil construction, the manufacture of household appliances and several other sectors.

A general equilibrium model is a mathematical representation of this network of relationships. It allows you to simulate what happens when there is a change at some point in the economy, and monitor how this change propagates through production chains.

General equilibrium models are particularly suited to capturing these cascading effects (Acemoglu et al. 2012; Carvalho, 2014), (Carvalho and Tahbaz-Salehi, 2019). It is possible, for example, that a relatively small shock in a very connected sector (which influences several production chains) could have amplified effects throughout the economy. An isolated sectoral analysis could estimate the direct impact of lost productivity in the oil refining industry, but it cannot

would capture how this loss impacts fuel prices, which in turn impacts transportation costs, which impacts the competitiveness of virtually every other sector of the economy.

Looking at our case, let’s imagine that law firms lose access to personalized AI tools and start using generic versions. What could happen?

• Professionals may spend more time reviewing contracts and reformulating analyses, and this could lead to an increase in the amount of fees charged;
• Their clients – banks and construction companies, industries – would receive opinions more slowly and at higher costs;
• Banks, in turn, could take longer to approve financing, delaying construction projects, and construction companies would purchase materials later, affecting cement and steel suppliers;

and so, no matter how small, the impact spreads: a loss of productivity in those who use AI to write and analyze information ends up spilling over into the entire production chain.

In summary, the general equilibrium model allows the study to go beyond the initial 1% shock to productivity. It is possible to estimate the final impact on the economy taking into account all the complex interactions between sectors, the effects of the tax system, and the specific characteristics of the Brazilian productive structure, as well as consumer responses.

The degree of exposure to the use of AI

How does the non-personalization of AI directly affect each sector of the Brazilian economy? It would be incorrect to measure the macroeconomic effect by assuming that all occupations within each sector have the same degree of exposure to these technologies.

To construct this measure of sectoral exposure, we combined three different sources of information:

• we use the exposure indexes by tasks developed by Felton, Raj and Seamans (2021), which measure how different types of tasks can be affected by AI;
• We combine this data with the exposure rates by work occupation de Cazzaniga et al. (2024), who estimate the proportion of tasks in each profession that could be impacted byLarge Language Models (LLMs); and, finally,
• we use the data from the National Household Sample Survey (PNAD) from IBGE, referring to the last quarter of 2024, to determine the weight of each occupation in each sector of the Brazilian economy.

EXPLAINING IN PRACTICE

Each occupation (profession) has a different level of exposure to AI, in addition, the tasks performed in each occupation also have different levels of exposure.

Let’s think about two occupations: journalists and doctors. For a journalist, tasks such as reviewing texts and transcribing interviews have a high exposure to AI

(i.e., AI can replace this task), while building a trusting relationship with sources and interpreting political nuances have less exposure (i.e., AI can hardly replace it). For doctors, tasks such as reading a simple imaging test or organizing medical records are highly exposed, while talking to the patient and deciding treatments in complex situations have much less exposure.

In this study, what we did was combine exposure rates by task with exposure rates by occupation, seeking a more precise look at the reality of professions. Along with this, we determined the weight of each occupation in the Brazilian economy based on IBGE data.

By combining exposure indices by occupation with data on the occupational composition of each sector of the Brazilian economy, the study was able to calculate a sectoral exposure measure, which indicates how much of each sector’s economic activity could be affected by the productivity shock caused by the customization ban.

Figure 1 – Quantitative analysis diagram

Notes: Schematic representation of the quantitative analysis in the three layers: (i) definition of shocks in each sector – as a combination of loss in productivity in tasks (1% drop) and sectoral exposure composed of the types of occupation in each sector of economic activity; (ii) simulation of general equilibrium effects in each sector and aggregate effects.

Note that the general equilibrium simulation carried out in this study does not consider possible dynamic changes with which AI is adopted over time, which reinforces the conclusion that the results consist of conservative estimates about the impact of banning the customization of AI tools.

The complete methodology of this study is detailed in the Appendix.

THE REGULATORY LENSES OF ARTIFICIAL INTELLIGENCEDiscussions about AI rarely start from a single regulatory lens. Each field – competition, data protection, digital rights, economic development – ​​observes the same phenomenon from legitimate but distinct concerns. These approaches may seem divergent, but in practice they complement each other: each offers a specific instrument to advance important public policy objectives. By placing this study within this plurality of lenses, the objective is to recognize that the economic perspective – centered on productivity, efficiency and aggregate impact – does not intend to replace, but to dialogue with other fields.

Results

Aggregate Macroeconomic Impact

By introducing shocks adjusted for sector-specific exposure into the general equilibrium model, banning AI customization would lead to a 1.64% drop in Brazilian Gross Domestic Product over a ten-year horizon, when compared to a reference scenario where the technology can be fully used.

It is important to emphasize that the effect of the shock occurs over time – and not all at once. In other words, when calculating the trajectory of the counterfactual scenario, it was considered what the decrease in the year-on-year growth rate should be to generate a GDP of 1.64%

smaller in 2035. Figure 2 presents the comparison between the base scenario (Baseline) and the Counterfactual simulation.

Figure 2 – Scenarios for GDP over a ten-year horizon

Notes: The black curve represents the baseline scenario and the orange curve represents the scenario from the simulation with the general equilibrium model.

To put it into perspective, Brazilian GDP in 2024 was approximately R$11.7 trillion. A 1.64% drop in GDP ten years ahead represents an accumulated loss (that is, the loss of GDP in each of the years of the simulation until 2035, brought to present value and added together) of R$855 billion.

It is important to understand why this 1.64% impact is greater than the initial 1% productivity shock: the model captures cascading effects through production chains. When a sector loses productivity, it not only produces less,

but it also makes their products more expensive, which affects all sectors that depend on these products as inputs. These sectors, in turn, also become less competitive, and the effect spreads.

Sector Impacts

	THE MOST AND LEAST IMPACTED SECTORS	dVA
1st	Manufacture of tobacco products	-1,913
2nd	Oil refining and coke plants	-1,897
3rd	Slaughter and meat products, including dairy and fish products	-1,886
4th	Biofuel manufacturing	-1,883
5th	Sugar manufacturing and refining	-1,873
6th	Agriculture, including agricultural and post-harvest support	-1,871
7th	Other food products	-1,856
8th	Livestock, including support for livestock farming	-1,849
Gº	Oil and gas extraction, including support activities	-1,821
10th	Electricity, natural gas and other utilities	-1,813
···	···	···
57th	Production of clothing items and accessories	-1,584
58th	Accommodation	-1,572
5Gº	Wholesale and retail trade	-1,564
60º	Artistic, creative and performance activities	-1,562
61st	Other administrative activities and complementary services	-1,561
62nd	Financial intermediation, insurance and supplementary pension	-1,554
63rd	Print-integrated publishing and editing	-1,549
64th	Private healthcare	-1,495
65th	Development of systems and other information services	-1,477
66th	Private education	-1,368

The sectoral analysis reveals that the impact would not be distributed evenly across the economy – and this reveals a pattern related to the Brazilian productive structure.

Sectors that appear at the top of the ranking are not necessarily those that use AI most directly, but rather those that are more deeply integrated with other sectors of the economy through input-output relationships, that is, the interdependencies between economic sectors that manifest themselves both through the use of inputs produced by other sectors and by offering products that will serve as inputs for others. Oil refining, the second most affected, supplies fuel and

raw materials for petrochemicals, the plastics industry and the entire transport chain.

When these sectors lose productivity due to bans on AI personalization, they reduce their demand for refined petroleum products, amplifying the impact on refineries. Likewise, agriculture provides essential inputs for the food industry, biofuels and various export products. As these purchasing sectors face higher costs and reduced production, demand for agricultural products also falls, multiplying the initial effect.

This dynamic of propagation through production chains explains why highly integrated sectors suffer disproportionately greater impacts: they not only face their own productivity losses, but also absorb the indirect effects of reduced demand from all sectors with which they relate.

On the other hand, it may seem contradictory that sectors such as private education, private healthcare or systems development are among the least affected. But this is explained by the production structure and the nature of the inputs:

• In these sectors, added value depends largely on the quality of decisions and interaction between people and knowledge specialized. Therefore, even if AI-based personalization is restricted, there is room for human adjustments that compensate for some of the loss in efficiency.
• Furthermore, these sectors tend to have relatively inelastic demand – health, education and financial services do not suffer sudden drops in consumption due to technological variations. The impact, therefore, is more diffuse and gradual, reflecting greater structural resilience to digital productivity shocks.
• In the case of the systems development sector, there is a paradox: although AI personalizes digital products, restricting customization can generate additional demand for adaptation, parameterization and technical support services, which would sustain the added value of the sector.

Looking again at the most harmed sectors, we see that these are highly integrated into standardized industrial chains and rely heavily on marginal efficiency gains. This means that:

• Even small productivity losses ripple throughout the supply chain. In these activities, AI personalization plays an increasing role in quality control, demand forecasting, logistics and energy optimization.
• When this layer of optimization is removed, the impact is multiplied: it affects costs, waste and deadlines, directly reducing the added value.
• Furthermore, These are export sectors with narrow margins, in which the international price limits the transfer of costs. Thus, any drop in efficiency reduces external competitiveness, magnifying the initial negative impact.

In summary, the results show that sectors based on human capital and with stable demand better absorb negative shocks from AI personalization, while sectors based on operational efficiency and production scale feel the impact more.

Analysis and Comments

The Transversality of AI and Personalization themes

The results of this study show that the Personalized Artificial Intelligence is not a topic restricted to large technology companies. AI crosses practically all productive sectors of the economy, from heavy industry to services. This

Diversity reveals that the debate on AI should not be treated only as a question of “digital platforms”, but also as a structural challenge of economic efficiency and national competitiveness.

Sectors traditionally seen as distant from the digital universe – such as oil and gas, agriculture – increasingly depend on AI systems to optimize processes, control quality, perform predictive maintenance and analyze large volumes of operational data. Furthermore, they are more affected by systemic impacts on the economy, precisely because of their centrality in production chains.

What could be impacted is not just the performance of technology companies, butall production chains in Brazil. AI regulation, therefore, must be understood as a broad development policy – ​​not just as a digital policy.

Strategic Implications

The results of this study do not just describe a loss of productivity. They reveal that discussing personalization – especially when we talk about personal data – is also discussing innovation and social well-being. The evidence in this report shows that the use of data in AI needs to be treated not just as a risk to be contained, but also as productive infrastructure that requires calibrated regulation and

incentives for responsible adoption. Based on the possibilities of using AI resulting from personalization, productivity gains are manifested not only across sectors, but propagate through complex production chains and can generate benefits for the economy as a whole.

The challenge now is to translate these discoveries into concrete actions and, to do this, it is necessary to guide different audiences with different, but complementary, messages.

The Brazilian regulatory agenda is experiencing a decisive moment. The debate about the Bill 2338/23 (which regulates the use of AI), developments and interpretations of GDPR in the courts and the ANPD, the approval of the ECA Digital and Bill 4675/25 (which regulates competition in digital markets) form a set of rules that seek to protect rights and preserve fundamental guarantees, but which can also generate social and economic costs greater than the expected benefits.

It is important that the regulator considers involving other sectors, especially industry and agriculture, in these discussions. Embed regulatory impact analyzes (AIRs) before proposing restrictions on the customization of AI systems and integrate productivity and economic efficiency datain debates about regulation are some of the practical recommendations that emerge from this study, which also reinforces the importance of using national empirical evidence as a reference in public consultations on AI.

For the private sector, the results reinforce that personalized AI is a factor of productivity and social legitimacy. Companies that use it

transparent, responsible and adjusted to the Brazilian context tend to capture gains in efficiency, reputation and more favorable regulation. On the other hand, ignoring the public debate about personal data, AI and personalization can generate image risks, reactive regulation and loss of international competitiveness.

It is important that companies from different sectors actively participate in the debate on digital regulation, participating in public consultations and sectoral coalitions that defend evidence-based policies, and that address the AI governance as part of the business strategy, not as an isolated technical topic.

Integration between Economics and Regulation

Although we have had more than a decade of robust discussions about data protection in the Brazilian regulatory scenario, economic studies have rarely been carried out in this field – which creates a false impression of tension between the economic evidence and the regulatory narrative.

These are not conflicting issues: are two legitimate ways of understanding technological progress. While the economy observes the efficiency and expansion of production and national income, among other factors, regulation is based on uncertainty and the need for protection in the face of the unknown. In the case of AI, this difference is

expands, especially when productivity gains are immediate and measurable, while risks – such as discrimination and reduced power of choice – are diffuse and long-term. Treating these two dimensions as incompatible is impoverishing the debate.

The quantitative results of this study, by suggesting that restrictions on AI personalization could reduce Brazilian GDP by 1.6% – compared to the scenario in which personalization is widely adopted – do not mean that regulation should be discarded, but that needs to be calibrated. Effective public policy is not one that only eliminates risk, but one that maintains space for regulatory learning – adjusting standards as empirical evidence accumulates. In this sense, the integration between economy and regulation requires a new type of institutional dialogue, which can point to a path of maturation in the Brazilian debate on AI.

By recognizing that economic gains and regulatory caution do not cancel each other out, but reinforce each other, the country can move towards a governance model thatuses economic evidence as a social protection instrument, and not as a counterpoint. Regulation informed by productivity and competitiveness data does not weaken individual rights: on the contrary, strengthens its sustainability by ensuring that public policies can be financed, tested and improved in a dynamic economy.

Conclusion

This study offered an unprecedented contribution by quantifying the macroeconomic impact of customizing Artificial Intelligence systems. By presenting the impacts on the Brazilian economy resulting from the ban on personalization (GDP 1.6% lower compared to the scenario of widespread use of personalization), the results provide a concrete empirical basis for a debate that, until now, has been conducted without a robust empirical basis.

This evidence is not intended to end the discussion, but to qualify it, introducing data that helps to measure the economic cost of different regulatory choices. More than a measurement exercise, the study proposes a methodological shift: understand AI not just as a technology, but as a vector of productivity across multiple sectors.

We believe that the findings can stimulate a more informed public debate, based on evidence, valuing public policies that reconcile development and data protection as parts of the same agenda.

Direction for Future Studies

Based on the results and discussions of this work, as well as its methodological limitations, we highlight the following directions for new studies, which can continue and improve this research:

• Empirical studies on social and cognitive impacts. Suggestion of controlled experiments and surveys to measure personalization effects on attention, well-being, and behavior;
• Distributional analyzes. Exploration of how the impacts of personalization (or its lack) affect employment, inequality and innovation across different sectors and social groups.
• International benchmarking. Comparative mapping of how different jurisdictions regulate different types of personalization, highlighting good practices and flaws that can inform the Brazilian debate.
• Custom AI Industry Models. Develop sector-specific econometric studies (such as agriculture, energy, and financial services) to identify variations in productivity gains and regulatory sensitivity to AI personalization.

AI Experimental Regulation

Test models regulatory sandbox for the personalization of AI systems, allowing to empirically measure impacts, risks and benefits before adopting general rules.

• Trust, Transparency and Social Acceptance. Conduct empirical research with users and consumers to understand how perceptions of transparency and control affect acceptance of AI personalization.

References

ACEMOGLU, D., C Azar, P. D. Endogenous production networks. Econometrica, v.88, n.1, p.33-82, 2020.

ACEMOGLU, D., et al. The Network Origins of Aggregate Fluctuations. Econometrica, v.80, n.5, p.1977-2016, 2012.

BAQAEE, D. R., Farhi, E., C Sangani, Kunal. The Darwinian Returns to Scale. The Review of Economic Studies, v.91, n.3, p.1373–1405, may. 2024.

BLOOM, Jan. Personalization – A Taxonomy. In: CHI2000: Human Factors in Computing Systems, 2000, The Hague. Student papers. Netherlands: 2000

BITRA, Sumar Kumar. The convergence of generative AI and hyper-personalization: Transforming customer experience at scale. World Journal of Advanced Research and Reviews, v.26, n.2, p.669-678, 2025.

Available at: <https://journalwjarr.com/sites/default/ files/fulltext_pdf/WJARR-2025-1648.pdf>. Accessed on: 22 Oct. 2025.

BRYNJOLFSSON, E., Li, D., C Raymond, L. Generative AI and Firm Value. NBER Working Paper, 2025.

CARVALHO, V.M. From Micro to Macro via Production Networks. Journal of Economic Perspectives, v.28, n.4, p.23-48, 2014.

CARVALHO, V. M., C Tahbaz-Salehi, A.Production Networks: A Primer. Annual Review of Economics, v.11, p.635-662, 2019.

CAZZANIGA, M., et al. Een-AI: Artificial Intelligence and the Future of Work. IMF Staff Discussion Note, 2024.

CHANDRA, Shobhana, et al. Personalization in personalized marketing: Trends and ways forward. Psychology C marketing, v. 39, no. 8, p. 1529–1562, 2022. Available at: <https://onlinelibrary.wiley.com/doi/ full/10.1002/mar.21670>. Accessed on: 22 Oct. 2025.

DELALIBERA, Bruno R., et. al.Tax reforms and network effects. Journal of Economic Dynamics and Control, V.163, 2024.

DELL’ACQUA, F., et al. Navigating the Jagged Technological Frontier: Field Experimental Evidence of the Effects of AI on Knowledge Worker Productivity and Quality. Harvard Business School Working Paper, 2023.

DILLON, E. W., et al. Shifting work patterns with generative ai. National Bureau of Economic Research, May. 2025.

ELOUNDOU, T., et al. GPTs are GPTs: An Early Look at the Labor Market Impact Potential of Large Language Models. arXiv preprint, 2024.

FELTON, R., Raj, M., C Seamans, R.. Occupational, industry, and geographic exposure to artificial intelligence: A novel dataset and its potential uses. Strategic Management Journal, v.44, n.1, p.1-28, 2021.

GADHVI, R., et al. AdaptAI: A Personalized Solution to Sense Your Stress, Fix Your Mess, and Boost

Productivity. In: Proceedings of the Extended Abstracts of the CHI Conference on Human Factors in Computing Systems, p.1-12, April. 2025.

GRAY, Daniel Berrick Stacey. Concepts in AI Governance: Personality vs. Personalization. Available at: <https://fpf.org/wp-content/ uploads/2025/09/Concepts-in-AI-eovernance_-Personality-vs.-Personalization.pdf>. Accessed on: 22 Oct. 2025.

HANDA, K., et al. Which economic tasks are performed with ai? evidence from millions of claude conversations. arXiv preprint arXiv, 2025.

KINDER, M., de Souza Briggs, X., Liu, S., C Muro, M. Generative AI, the American worker, and the future of work. 2024.

KWON, Kwiseok; KIM, Cookhwan. How to design personalization in a context of customer retention: Who personalizes what and to what extent? Electronic commerce research and applications, vol. 11, no. 2, p. 101–116, 2012. Available at: <https://www.sciencedirect. com/science/article/abs/pii/S1567422311000238>.

Accessed on: 22 Oct. 2025.

LEROY, S. Why is it so hard to do my work? The challenge of attention residue when switching between work tasks. Organizational Behavior and Human Decision Processes, v.109, n.2, p.168-181, 2009.

MARK, G., Eudith, D., C Klocke, U.. The cost of intermittent work: more speed and stress. Proceedings of the SIGCHI conference on Human Factors in Computing Systems, p.107-110, 2008.

NISAR, Tahir. The personalization economy: how is AI affecting businesses and markets?.

Economics Observatory. Available at: <https://www. economicsobservatory.com/the-personalization-economy-how-is-ai-affecting-businesses-and-markets>. Accessed on: 22 Oct. 2025.

NOY, S., C Zhang, W.. Experimental Evidence on the Productivity Effects of Generative Artificial

Intelligence. Science, v.381 n.6659, p.778-783, 2023.

OSWALD, A. J., Proto, E., C Sgroi, D.. Happiness and Productivity. Journal of Labor Economics, v.33, n.4, p. 789-822, 2015.

SHAFIK, Md Rehman. AI and Personalization in Digital Marketing.International Journal of Progressive Research in Engineering Management and Science (IJPREMS) (Int Peer Reviewed Journal). v.5, n.6, p. 1294-1307, jun. 2025.

VESANEN, Jari. What is personalization? A conceptual framework. European journal of marketing, vol. 41, no.

5/6, p. 409–418, 2007. Available at: <https://2024.

sci-hub.se/3591/1a1be69c794c07e95c755d69c19f6f10/vesanen2007.pdf>.

Annex: Impact ranking in sectors

	RANKING: IMPACT ON SECTORS	dVA
1st	Manufacture of tobacco products	-1,913
2nd	Oil refining and coke plants	-1,897
3rd	Slaughter and meat products, including dairy and fish products	-1,886
4th	Biofuel manufacturing	-1,883
5th	Sugar manufacturing and refining	-1,873
6th	Agriculture, including agricultural and post-harvest support	-1,871
7th	Other food products	-1,856
8th	Livestock, including support for livestock farming	-1,849
Gº	Oil and gas extraction, including support activities	-1,821
10th	Electricity, natural gas and other utilities	-1,813
11th	Manufacture of organic and inorganic chemicals, resins and elastomers	-1,811
12th	Forestry production; fishing and aquaculture	-1,781
13th	Extraction of iron ore, including processing and agglomeration	-1,775
14th	Extraction of non-ferrous metallic minerals, including processing	-1,772
15th	Manufacture of pesticides, disinfectants, paints and various chemicals	-1,767
16th	Manufacture of cleaning, cosmetics/perfumery and personal hygiene products	-1,764
17th	Metallurgy of non-ferous metals and metal casting	-1,763
18th	Beverage manufacturing	-1,759
1Gº	Manufacture of cars, trucks and buses, except parts	-1,734
20th	Production of pig iron/ferroalloys, steelmaking and seamless steel tubes	-1,732

	RANKING: IMPACT ON SECTORS	dVA
21st	Land transport	-1,720
22nd	Manufacture of pulp, paper and paper products	-1,718
23rd	Extraction of mineral coal and non-metallic minerals	-1,717
24th	Manufacture of rubber and plastic products	-1,711
25th	Water transport	-1,709
26th	Manufacture of products from non-metallic minerals	-1,701
27th	Manufacture of electrical machinery and equipment	-1,695
28th	Manufacture of metal products, except machinery and equipment	-1,694
2Gº	Construction	-1,686
30th	Manufacture of parts and accessories for motor vehicles	-1,682
31st	Maintenance, repair and installation of machines and equipment	-1,678
32nd	Public education	-1,677
33rd	Public health	-1,677
34th	Public administration, defense and social security	-1,677
35th	Manufacturing of wood products	-1,671
36th	Manufacturing of textile products	-1,669
37th	Manufacture of shoes and leather goods	-1,669
38th	Manufacture of other transport equipment, except motor vehicles	-1,665
3Gº	Food	-1,662
40º	Manufacture of computer equipment, electronic and optical products	-1,661
41st	Real estate activities	-1,661
42nd	Telecommunications	-1,655
43rd	Air transport	-1,655

	RANKING: IMPACT ON SECTORS	dVA
44th	Other professional, scientific and technical activities	-1,644
45th	Television, radio, cinema and sound and image recording/editing activities	-1,640
46th	Legal, accounting, consultancy and company headquarters activities	-1,639
47th	Non-real estate rentals and intellectual property asset management	-1,637
48th	Manufacture of machines and mechanical equipment	-1,627
4Gº	Surveillance, security and investigation activities	-1,626
50th	Storage, auxiliary transport and mail activities	-1,625
51st	Manufacture of pharmochemical and pharmaceutical products	-1,624
52nd	Manufacture of furniture and products for various industries	-1,615
53rd	Water, sewage and waste management	-1,613
54th	Architecture, engineering, technical testing/analysis and P C D services	-1,609
55th	Printing and playing back recordings	-1,597
56th	Membership Organizations and Other Personal Services	-1,587
57th	Production of clothing items and accessories	-1,584
58th	Accommodation	-1,572
5Gº	Wholesale and retail trade	-1,564
60º	Artistic, creative and performance activities	-1,562
61st	Other administrative activities and complementary services	-1,561
62nd	Financial intermediation, insurance and supplementary pension	-1,554
63rd	Print-integrated publishing and editing	-1,549
64th	Private healthcare	-1,495
65th	Development of systems and other information services	-1,477
66th	Private education	-1,368
			2C

Reglab Methodology Annex

AUTHORITY: João Ricardo Costa Filho and Pedro Henrique Ramos

Title	The Economic Dimension of Personalization: Measuring the Impact of Personalized Artificial Intelligence on Brazilian GDP
Research Question Methodology Summary Data Collection Data Analysis	What is the macroeconomic impact of banning the customization of artificial intelligence tools on the Brazilian economy?
Research Question Methodology Summary Data Collection Data Analysis	Calculation of sectoral exposure to Artificial Intelligence; General equilibrium control model calibrated for the Brazilian economy. Base scenario projection for Brazil’s GDP over the next 10 years. Calculation of the present value of the impact of aggregate productivity loss.
Research Question Methodology Summary Data Collection Data Analysis	Desk research of secondary data, as follows: Exposure index by activity (Felton, Raj, and Seamans, 2021). Exposure index by occupation (Cazzaniga et al., 2024) American Occupation Category (SOC) International Occupation Category (ISCO) Brazilian occupation category (CBO) Microdata from the IBGE National Household Sample Survey (PNAD) Last quarter of 2024; 2024 GDP: R$11.7 trillion (Source: IBGE, SCNT) Real GDP growth rates projected in the Focus Report (BCB, 12/Sep/25).
Research Question Methodology Summary Data Collection Data Analysis	The model used in this study is based on the work of Delalibera (2024). This model was chosen because it incorporates important characteristics of the Brazilian economy that more generic models do not have. First, it includes a productive public sector, recognizing that the Brazilian State is not just a regulator or consumer, but also a producer in areas such as energy and services. Second, the model incorporates a complex tax system that considers the different tax rates that apply to different sectors and products. Third, the model recognizes that many markets do not operate in perfect competition, but rather in imperfect competition regimes, where companies have market power and can decide on the pricing of their products. This characteristic, inspired by the work of Baqaee and Farhi (2020) and Acemoglu and Azar (2020), is important because it affects how productivity shocks translate into changes in prices and quantities. The quantity and definition of sectors are derived from the input-output matrix calculated by the Brazilian Institute of Geography and Statistics (IBGE). Furthermore, the following references were used for data analysis and calculations: Exposure index by activity (Felton, Raj, and Seamans, 2021). Exposure index by occupation (Cazzaniga et al., 2024) Weight of each occupation in the Brazilian economy: 1st step: compatibility of occupation categories: American occupation category (SOC) -> International occupation category (ISCO) -> Brazilian occupation category (CBO) 2nd step: National Household Sample Survey (PNAD) by IBGE Last quarter of 2024; Microdata: weight of individuals, with an occupation (CBO) linked to a CNAE; Calculation of the weight of each occupation in each CNAE; Sum of (Exposure per occupation x weight of each occupation in each CNAE) = average exposure for each sector.

Data Analysis Bias Reduction Procedures Other Methodological Limitations Use of Software	Construction of the reference scenario: 2024 GDP: R$11.7 trillion (Source: IBGE, SCNT) Real GDP growth rates projected in the Focus Report (BCB, 12/Sep/25) For 2029 and 2030 the rates from 2028 were repeated; Discount rate: Selic rate projected in the Focus Report; For 2029 and 2030 the rates from 2028 were repeated; Inflation rate (IPCA) projected in the Focus Report; For 2029 and 2030 the rates from 2028 were repeated; It is important to highlight that the functions that were not present in the International Standard Classification of Occupations (ISCO) were excluded and the remaining functions were reweighted. When constructing the reference scenario, the GDP of the year 2024 and the real growth rates projected by the median of the Focus Report respondents were considered, from the Central Bank of Brazil, on October 10, 2025. For the years 2029 onwards, the same growth rate projected for the year 2028 (2%) was considered. Similarly, projections for the Selic rate and IPCA were used to calculate the timeless discount factor, that is, the factor used to calculate the present value of potential GDP losses, resulting from the non-personalization scenario, which may occur in the future.
Data Analysis Bias Reduction Procedures Other Methodological Limitations Use of Software	To reduce bias, empirical analysis references widely consolidated in the literature were used. Furthermore, the methodological approach was discussed and evaluated internally on two occasions so that suggestions and criticisms could be incorporated into the work before the analysis was carried out. Other procedures adopted include: Double Validation in Critical Stages: for the “Analysis and Comments” section, the two authors reviewed the text independently. In cases of disagreement, a third person was called to arbitrate and reach consensus. Recording and Methodological Transparency: we kept detailed records of all versions of files and research, preserving history and allowing for a more systematic review.
Data Analysis Bias Reduction Procedures Other Methodological Limitations Use of Software	Reliance on Open Access Sources: The study relied significantly on searches conducted in open access databases and academic journals. Dependence on these sources may restrict the scope of the analysis, considering that relevant materials present in restricted access or specialized databases may not have been considered, which may compromise the completeness and depth of the text presented.
Data Analysis Bias Reduction Procedures Other Methodological Limitations Use of Software	Software Use in Research MS Office suite Text editing, spreadsheets and graphics Adobe C Suite Layout and finalization of graphics and illustrations Grammar review (spelling, grammar, Google search eemini synonyms), language adaptation and pre-prepared excerpts R Compatibility of different databases and estimation of sectoral exposure to AI. Matlab Simulation of the general equilibrium model. Chat9PT Assistance in programming simulations. Manus Writing development.

Ethical Guidelines

This research was funded by Facebook Serviços Online do Brasil Ltda (“Meta”). To guarantee the integrity of this work, the authors developed, conducted and analyzed the study independently, without any contribution or interference from the company, which also did not influence or interfere in the interpretation of the results. The authors maintain full professional independence and responsibility for the content and conclusions of this work. Respect for Privacy and Confidentiality: The data used is in the public domain and was obtained from accessible sources, without violating the privacy or confidentiality of any individual or institution. Responsible Use of Public Data: Although the data analyzed is public, its use was made in a responsible and ethical manner, with the exclusive purpose of academic research. Non-discrimination and Respect for Diversity. The research was conducted in a way that respects diversity and avoids any form of discrimination.