- Inteligência Artificial
- Futuro do Trabalho
- Impacto da IA
- Mercado de Trabalho
- Requalificação Profissional
- IA no Desenvolvimento
AI and the Future of Work
Tasks Change Before Job Titles
"Will AI take my job?" The short answer, in most cases, is no. But it will change what you do in it, and some roles will need fewer people. Both things are true at the same time, and most of the debate about AI and the future of work picks only one: either "no job is safe" or "AI only helps". In this post I separate tasks from occupations using data from the ILO, Stanford, and the World Economic Forum. Then I show who is most exposed, share what changed in my own work as a developer, and close with an exercise you can apply to your role.
Tasks, Not Occupations: What AI Exposure Really Means
Almost no occupation is a single task. A support agent answers questions, but also calms down angry customers, handles exceptions, and notices when a problem is bigger than it looks. A developer writes code, but also understands the problem, picks the architecture, and decides what not to build. AI takes some of those tasks, not the whole package.
That is why the right question is not "will my occupation disappear?" but "which tasks in my occupation are exposed?". Being exposed to AI means part of your work can be done or sped up by it. It does not mean being replaced.
Numbers from the International Labour Organization show this difference clearly. In the 2025 update of its exposure index, 1 in 4 workers worldwide is in an occupation with some exposure to generative AI. But only 3.3% of global employment sits in the highest exposure level. The ILO's own conclusion:
"As most occupations consist of tasks that require human input, transformation of jobs is the most likely impact of GenAI."
In other words: since most occupations include tasks that need a human, changing how work is done is the most likely outcome, not wiping jobs out.
It changes, it creates, and it also shrinks
"Transformation" does not mean "nothing happens". I see three movements happening together.
- Tasks change. The same role gets a different day-to-day. This is the most common case.
- New roles appear. My current position is an example: since June 2026 I have led the AI Enablement area at InBot, a role that did not exist at the company until recently.
- Some roles do shrink. They do not vanish, but they need fewer people.
The World Economic Forum's Future of Jobs Report 2025 puts numbers on both sides: by 2030, 170 million jobs are expected to be created and 92 million displaced, a net gain of 78 million. The overall balance is positive, but it is no comfort to those among the 92 million. That is why it is worth looking at who is closest to that side of the equation.
Who Is Most Exposed: Office Work, Customer Support, and Early Careers
Back in 2023, the ILO already pointed to clerical work as the most exposed: nearly a quarter of the tasks in those occupations are highly exposed to generative AI. The WEF lands in a similar place when it lists the fastest-shrinking roles: cashiers, administrative assistants, data entry clerks, telemarketers, and graphic designers.
In my view, four kinds of work are shrinking:
- Repetitive office tasks: triage, form filling, standard reports.
- Tier-1 support: frequently asked questions and simple tickets.
- Routine code: CRUD, boilerplate, translating code between languages.
- Generic content produced at volume: low-value text and images.
The last one already has data. A study of freelancers found that, after ChatGPT launched, writers lost about 2% of their jobs per month and 5.2% of their income. It is a short-term effect, measured in 2023, but it points in the expected direction.
Shrinking does not mean cutting everything, either. In customer service, Gartner predicts that half of the companies that cut staff because of AI will rehire by 2027. The simple task leaves human hands. The hard case still needs someone.
Juniors lose and gain at the same time
The data point that strikes me most is about people starting out. The "Canaries in the Coal Mine?" study from the Stanford Digital Economy Lab was updated in August 2026 with US payroll data. It shows that employment of 22 to 25-year-olds in the occupations most exposed to AI is 19% below where it would be had it kept pace with less-exposed occupations. Among experienced workers in those same roles, no comparable gap appears. Software development and customer service are among the examples.
Two details in the study matter a lot. First: the effect comes from less hiring of young people, not more layoffs. Second: it shows up where AI replaces tasks. Where it complements the work, employment stays stable or grows.
In Brazil, the pattern repeats. A study by Daniel Duque at FGV Ibre found negative effects on employment and income only among 18 to 29-year-olds more exposed to AI. For ages 30 and up, the effect is consistent with zero. The author's explanation sums it up: AI lowers the value of standardized tasks, exactly the ones that used to be the entry point to many careers.
These findings are still early and, on their own, do not prove causation. The Stanford authors themselves say AI is not the only factor: interest rates and earlier trends also play a role. Even so, the first data points in a clear direction.
Now, the other side. The NBER study of more than 5,000 support agents found that AI raised productivity by 14% on average and by 34% among novice and less-skilled workers. For the most experienced, the effect was minimal. In other words, AI shortens the learning curve for people who are starting out.
That is why I think juniors lose and gain at the same time. Those who use AI to learn get a shortcut. But entry-level openings for simple tasks are shrinking, because those tasks were exactly the ones AI learned to do. In software, Indeed Hiring Lab shows the scale of it: only 4.5% of software development postings in the US are entry-level, and 71% of the recent recovery in postings came from senior roles.
Software development shows up in almost all of this data. It is also the profession I know from the inside.
The Case of Software Development: From Writing Code to Orchestrating and Validating
My work has changed a lot in recent years. I used to write code. Today I orchestrate: I think through the architecture, define the requirements, break the work into tasks, create agents with clear limits, and validate what the AI delivers. I described this process in detail in my AI development workflow.
That does not mean I need to know less. I need to know more, because I have to understand and validate code I did not write. And I am not an exception. In the 2025 Stack Overflow survey, 84% of developers use or plan to use AI. At the same time, trust in the accuracy of its output dropped from 43% to 33% in one year. The top frustration, cited by 66%, is code that is "almost right, but not quite".
Google's DORA 2025 report sums it up in one sentence: AI's primary role in software development is that of an amplifier. It magnifies what a team does well, and also what it does badly.
What makes a good developer now
If AI writes much of the code, a developer's value shifts to what comes before and after it. For me, it is four things together:
- Architecture and requirements: knowing what to build and why, before asking for a single line of code.
- Validating what AI generates: tests, review, and security. This is where TDD comes into my work: the test defines the behavior, and the code is what adapts to it.
- Orchestrating agents: breaking the work into small tasks and giving each agent clear limits on what it can and cannot do.
- Domain knowledge: understanding the business and the customer better than any tool.
The risks: vibe coding today, lost foundations tomorrow
In the short term, the biggest risk is vibe coding without a foundation: people generating entire systems without understanding what was generated. It works until the first security problem shows up. Out of curiosity, I have looked into applications built this way and got full access to more than one.
In the long term, the risk is losing the foundations. People who never wrote code have a harder time reading code, and people who cannot read it cannot validate it. Anthropic ran an experiment with engineers, mostly juniors, learning a new library. Those who learned with AI scored 50% on a comprehension test, versus 67% for those who coded by hand, and the biggest gap was in debugging. The sample is small, and the company sells the product being studied. Even so, the authors' sentence sums up the risk:
"Productivity benefits may come at the cost of skills necessary to validate AI-written code."
In other words: the productivity gain may cost exactly the skills you need to check AI-written code. The same study offers a useful clue: those who used AI to ask questions about the concepts, and not just to generate code, learned more.
AI is not a tool, it is a technology
Many people talk about the risk of "depending on the tool". I see it differently. AI is not a tool, it is a technology. We all learned to do math by hand at school, but no workplace requires us to do that on the job. They all require us to use a calculator, because it automates the process.
Depending on the calculator was never the problem. The problem would be not knowing what multiplication is and, because of that, not noticing when the result makes no sense. It is the same with AI: using it will be mandatory. What sets people apart is knowing the fundamentals well enough to spot the mistake.
And this logic does not apply only to developers.
Future Skills: What Becomes More Valuable (and What Becomes Less)
The WEF estimates that 39% of workers' core skills will change or become outdated by 2030. A word of caution on that number: it has been falling (it was 57% in 2020 and 44% in 2023), partly because more people have received training. It is not an acceleration. It is a large and ongoing shift. At the top of the report's list is analytical thinking, considered essential by 7 out of 10 companies.
In my view, four skills become more valuable in any field:
- Critical thinking: questioning the AI's answer and knowing when it is wrong.
- Domain knowledge: understanding the business and the customer. AI knows a lot in general and nothing about the exception your team handles every week.
- Process design: before automating, asking whether the problem is speed or the process itself. AI speeds up what exists. If the process is wrong, it speeds up the mistake.
- Writing and clarity: those who write well, ask well. Explaining precisely what you want has become a work skill.
What becomes less valuable is memorizing and executing: knowing by heart the commands, formulas, and steps that AI executes better and faster. Knowing how to execute is still useful, it just stopped being the differentiator.
The market already pays for this shift. According to PwC, job postings that ask for AI skills pay 62% more on average. And the entry-level postings most exposed to AI now ask far more often for senior-level skills, such as leadership and creativity. PwC sells AI consulting, so read it with care, but the data matches Stanford's: the entry door does not disappear, it gets more demanding.
If skills change this much, one question remains: who pays for it?
Reskilling: Whose Responsibility Is It
For me, the responsibility is shared. The individual seeks it out, the company trains, and the public sector guarantees the basics. When one of the three fails, the people who pay are those with the fewest resources to adapt on their own.
The individual: seek it out
Nobody will map the tasks of your role better than you. Waiting for your company or the government to act before you learn is risky, especially early in your career, where the data shows the biggest impact. Seeking it out does not have to mean an expensive course. It can mean using AI in your own work, every day, with a critical eye.
The company: train
On paper, companies already agree: 85% of the employers surveyed by the WEF say they will prioritize reskilling. In practice, training is not keeping up. According to Microsoft and LinkedIn's 2024 Work Trend Index, only 39% of AI users had received any training from their company. A 2025 survey by BCG found only a third of employees saying they were well trained. And the WEF itself estimates that, out of every 100 workers, 11 will likely not receive the training they need by 2030.
I argue that training is smarter than cutting, and I wrote about how this works inside a company in AI governance in practice.
The public sector: public policy and social protection
This is the part that usually gets left out of the debate. Individuals and companies solve a lot, but not everything. Someone who loses an entry-level job at 23, in a role that shrank, needs accessible training and protection during the transition. The ILO has said this plainly since 2023: workers' voice, skills training, and adequate social protection will be key to managing this change.
In Brazil, there are two developments worth following:
- Bill PL 2338/2023, which regulates AI, was approved by the Senate in December 2024. The text has a dedicated section on protecting work: it calls for guidelines to reduce negative impacts on workers, "especially the risks of job displacement", and for public programs for reskilling and supporting workers affected by AI. But it is a bill, not a law. As of October 2026, it was still awaiting the rapporteur's report in the Chamber of Deputies.
- The Brazilian AI Plan sets aside about R$ 23 billion through 2028. Of that total, R$ 500 million goes to the AI training, qualification, and reskilling program. It is a start, but it is just over 2% of the plan.
Brazil currently has no social protection policy specifically for people who lose their jobs because of AI. Maybe it does not even need to be specific. But if early careers are the most fragile point, that is where public training and protection during the transition make the biggest difference.
Do the Exercise with Your Own Role
Data helps you understand the landscape, but the decision is about your work. So here is a simple exercise:
- List the tasks in your week. Not your job title: the tasks. Answering tickets, reviewing contracts, building reports, writing tests, serving customers.
- Sort each one into three groups. What AI already does well today. What AI amplifies, making you faster or better. And what still depends on you: judgment, relationships, domain knowledge, decisions.
- Look at the size of each group. If the first is the largest, your role is among those that shrink, and it is worth starting to shift your time to the other two. If the third is the largest, invest in using AI in the second to gain scale.
Conclusion
AI will not end most occupations, but it will not leave any of them the same. Tasks change first, new roles appear, and some roles start needing fewer people. The greatest pressure is on early careers, where the simple tasks that used to teach the craft are exactly the ones AI learned to do.
What protects you is not running from AI. It is what has always protected people: understanding the fundamentals, knowing the domain, thinking critically, and being able to explain what you want. Nobody does math by hand at work, but those who understand the math notice when the calculator got it wrong.
Do the three-column exercise with your role this week. And if you lead a team, do it with them: reskilling starts when someone stops and looks, task by task, at what is changing.
References
- ILO; Gmyrek, P.; Berg, J.; Kamiński, K. et al. "Generative AI and Jobs: A Refined Global Index of Occupational Exposure". ILO Working Paper 140, 2025.
- ILO; Gmyrek, P.; Berg, J.; Bescond, D. "Generative AI and Jobs: a Global Analysis of Potential Effects on Job Quantity and Quality". ILO Working Paper 96, 2023.
- ILO. "Generative AI likely to augment rather than destroy jobs", 2023.
- World Economic Forum. "The Future of Jobs Report 2025".
- Brynjolfsson, E.; Chandar, B.; Chen, R. "Canaries in the Coal Mine? Six Facts about the Recent Employment Effects of Artificial Intelligence". Stanford Digital Economy Lab, August 2026 version.
- Brynjolfsson, E.; Chandar, B.; Chen, R. "Canaries, Interest Rates, and Timing: More on Recent Drivers of Employment Changes for Young Workers". Stanford Digital Economy Lab, 2026.
- Duque, Daniel. "Inteligência artificial generativa e mercado de trabalho no Brasil: evidências iniciais sobre ocupação e renda". IBRE/FGV Blog, 2026.
- Brynjolfsson, E.; Li, D.; Raymond, L. "Generative AI at Work". NBER Working Paper 31161 / Quarterly Journal of Economics, 2025.
- WashU Olin. "Study: AI tools cause a decline in freelance work and income, at least in the short run" (Hui, Reshef and Zhou), 2023.
- iTnews Asia. "Half of firms that cut customer service staff due to AI will rehire by 2027" (Gartner), 2026.
- Aidala, F.; Puri, S. "The Labor Market Is Tilting Toward Seniority". Indeed Hiring Lab, 2026.
- Gallacher, G. "AI and Job Postings: From Destruction to Creation?". Indeed Hiring Lab, 2026.
- Stack Overflow. "2025 Developer Survey: AI".
- Google Cloud / DORA. "State of AI-assisted Software Development 2025".
- Anthropic. "How AI assistance impacts the formation of coding skills", 2026.
- PwC. "2026 Global AI Jobs Barometer" (press release), 2026.
- Microsoft and LinkedIn. "2024 Work Trend Index".
- BCG. "AI at Work: Momentum Builds, but Gaps Remain", 2025.
- Brazilian Chamber of Deputies. Legislative tracking of Bill PL 2338/2023.
- Brazilian Chamber of Deputies. Full text of Bill PL 2338/2023 (as approved by the Senate).
- MCTI. Presentation of the Brazilian Artificial Intelligence Plan to the Chamber's CCTCI committee, 2025.
- Mierzwa, Matheus. "My AI Development Workflow".
- Mierzwa, Matheus. "My TDD Journey".
- Mierzwa, Matheus. "AI Governance in Practice".