<AI Meets the Todero/>

The Todero Enters the Scene

In Colombia, Venezuela, and parts of Central America, a “todero” is a worker who does everything on her own. Think of the plumber who also writes the quotes, manages the materials, tracks inventory, handles customer complaints, and posts on Instagram.

Or the baker who runs the shop, the accounts, the supplier relationships, and the delivery. The todero does not fit neatly into any organisational chart because there is no chart, just a person and a business to keep alive.

The data suggests that this is not an anecdote. According to Atencio-De-Leon et al. (2025), in a matched sample of 101 occupations, the average Peruvian occupation activates 7.1 of nine skill families, compared with 3.6 in the U.S. O*NET benchmark; Peru shows a broader portfolio in 94 of the 101 occupations.

And in nearly 5.9 million job postings from Argentina, Brazil and Uruguay analysed by Carbajal et al. (2025), broad skill demand appears across the occupational hierarchy: from managers to elementary occupations, median breadth remains within a surprisingly narrow range.

The todero, in other words, is not a niche occupational category. It is the way work is organised across the occupational spectrum in Latin American labor markets. [4,5]

Bandiera et al. (2022) find a closely related pattern at scale in their Jobs of the World project: the organisation of labour differs systematically across development levels in ways that occupational labels alone do not capture. [6]

When an index assigns 'low exposure' to a plumber in Lima, it may be describing the task profile of a specialised plumber - not the broader task portfolio of the person who actually shows up for work every morning in the busy neighborhoods of Miraflores or Callao.

This has a direct implication for how we measure AI exposure. Many influential exposure measures rely on standardised occupational task descriptions that are then applied across labour markets. [1-3]

That is often a useful approximation, but it embeds a strong assumption: the same occupational label implies a broadly similar task bundle everywhere. If the division of labour differs systematically across countries, some of the tasks that matter most for AI may sit outside the imported occupational profile.

Low Specialisation, and “Good” Automation

Before asking which todero's daily activities are exposed to AI, it helps to understand why the todero exists in the first place, and what her existence says about stagnant productivity and weak growth.

The classical answer goes back to Adam Smith: the division of labour is limited by the extent of the market. In thin markets, workers absorb more functions because there is not enough demand to sustain specialists in each one.

Rosen (1983) formalised the logic for human capital: specialisation is not just a technical property; it is a market outcome. [7] Baumgardner (1988) showed the same logic empirically: in smaller markets, professionals cover broader domains because the market cannot sustain the same degree of division of labour. [8]

The todero, then, is an organisational response to a market with limited depth. Latin America's firm-size distribution makes this especially relevant: self-employment and very small firms account for a large share of employment, settings in which specialised support is comparatively scarce. [9]

That equilibrium carries a cost. A long-standing diagnosis of Latin America's development gap is that low efficiency - not only limited accumulation of capital and skills - accounts for a substantial part of its income shortfall.

Delayed technology adoption, misallocation, and organisational inefficiency within firms are among the underlying constraints. [10] If the way work is organised is part of the productivity problem, then changing the organisation of work can also be part of the solution.

This is where the familiar automation-versus-augmentation debate starts to look different. In a polyfunctional job, automating a surrounding task need not weaken the worker's core contribution. It can free scarce time and attention for the parts of the job where the worker has comparative advantage.

AI Enters the Scene

For a highly specialised worker, AI enters an already differentiated labor market structure. The lawyer may already have research assistants; the analyst may already have data teams and specialised software.

For the todero, the dynamic is structurally different. AI is often not competing with a specialist who already exists on the team - because that specialist does not exist.

It can supply writing, search, basic analysis, formal communication, planning, or administrative capacity that the thin market never provided.

Consider a baker in a mid-sized Latin American city. In the Peru-U.S. comparison, the difference between occupational labels and actual skill breadth can be dramatic.

The technical core - production, quality, the craft of baking - defines the occupation in a standard profile. But the Peruvian baker also activates organisational, cognitive, writing, project-management, customer-service, financial, people-management, and social skills.

And this is where the match becomes especially important. The additional tasks carried by workers in less-specialised labour markets are often precisely the kinds of activities in which AI is already being used.

Bick et al. (2026), for example, find substantial adoption in activities such as extracting technical information from documents, preparing reports and analysing data. [13]

Large-scale evidence on actual AI use points in the same direction: writing, searching, organising, synthesising and other cognitive and administrative activities feature prominently in observed use. [14,15]

In other words, the tasks that make the todero’s job broader are often the same tasks where AI is already proving useful.

This does not mean that AI will necessarily be adopted at scale in Latin America. Country-level usage data already reveal large and persistent gaps in AI diffusion across economies, even if different sources disagree on whether those gaps are currently widening or beginning to narrow.

The distance between potential exposure and effective use is large, and closing it requires a broad set of complementarities—from infrastructure and connectivity to appropriate technologies, worker skills and organisational capabilities.

It tells us this: if workers in less-specialised labour markets perform broader portfolios of tasks than conventional occupational profiles imply, then the surface of work that AI can potentially transform may also be broader than our standard maps suggest.

AI, Specialisation, and Productivity

Get back to the Peruvian baker. Hand some of those peripheral functions to a reliable AI assistant and she does not become a different kind of worker. She may become a more productive baker - and perhaps run a more capable business.

Gans and Goldfarb's (2026) focus mechanism shows how automating one task can release scarce attention for the remaining bottlenecks; Althoff and Reichardt's (2026) simplification mechanism shows how technology can lower the skill requirement of activities that previously required specialist support. [11,12]

For a todero, both can operate at once: AI can remove low-value overhead while making capabilities such as market research, product development, formal communication or basic analysis easier to access.

The result could be more time for the craft and the customer relationship, but also better products, more professional business practices, and access to customers and markets that were previously out of reach.

If these mechanisms hold, the implications extend beyond job exposure indices. If AI allows toderos and small firms to shed low-value surrounding tasks and concentrate human effort on higher-value activities, it can raise productivity within existing jobs and productive units - not only at the technological frontier, but throughout the distribution.

This is not the conventional story of a technology invented in rich countries slowly diffusing to poorer ones. It is a story about using AI to reorganise work where occupational variety is limited and organisational depth is scarce.

The potential gains on this margin may therefore be larger where the organisational gap is larger. That is precisely why the mechanism may be informative beyond Latin America for other Global South economies with similarly thin markets, small productive units, and limited specialisation.

Read from this perspective, Amodei's (2024) image of a 'country of geniuses in a datacenter' takes on a different - and potentially broader - labour-market meaning. [16]

Where specialised human support is scarce, those cognitive capabilities could become available to workers and small firms across the occupational spectrum. In that sense, transformative AI may be transformative not only at the frontier, but across a much larger share of the global labour market.

For economies that have struggled for decades to close productivity gaps, that also opens a potential new channel for catch-up growth.

Change the Lens to Redesign the Future

The opportunity is not to preserve existing jobs exactly as they are. Against the backdrop of technological anxiety, we are proposing a countercultural yet needed strategy: to foster automation.

Not any automation, but to automate the right things, deepen effective specialisation, and build stronger, more productive human bundles around what remains. The goal is not to defend the current division of labour. It is to use AI to redesign it.

AI opens a window to design a better future of work. To seize that opportunity, we need to bring AI closer to workers through inclusive technology: mobile-first interfaces, voice and image interaction, local language and domain vocabulary, transparent uncertainty, and simple ways to verify outputs and retain control.

Workers also need opportunities to build the capabilities required to identify what can be delegated, formulate useful requests, evaluate results, and reintegrate AI into real workflows.

This calls for situated AI training: learning AI not through generic prompting exercises (“chatgpt for dummies”), but inside the actual task bundles of an occupation.

A self-employed electrician, for example, might learn with her own workflow— preparing a quote, comparing materials, communicating with a client, checking regulations or organising a schedule—while also learning which decisions should remain hers and how to verify what the system produces.

At Sur Futuro, this is the direction we are beginning to explore: training organised around real occupations, real tasks and real problems, with AI as a tool embedded in work rather than as a skill taught in isolation.

That window also requires a new lens on the labour market. Imagine having comparable public data that could serve as O*NET-like satellites outside the United States, capturing how occupations are actually organised across different labour markets.

Better statistics are needed, for sure, but it’s not enough. We also need to go out there and get closer to the real world of work: to observe who does what, when and where, how tasks unfold over time, and how workers respond when conditions change.

The goal is simple: stop inferring work only from occupational labels and go observe how people actually work.

Work diaries can reconstruct the sequence of tasks people actually perform; experience-sampling methods can capture activities, interruptions and surrounding functions in real time; and AI itself can make it possible to collect, transcribe, classify and analyse this kind of granular information at a scale that was previously prohibitively costly.

And when the question is causal -or when we want to design interventions rather than simply describe work- we need more field experiments, including randomized controlled trials (RCTs), like the one we conducted with Cruces et al. (2026) [17].

Better maps, richer observation and more experimentation can help us discover which combinations of human and AI capabilities actually create value.

And that takes us back to Greenmarket Square. The workers who initially looked peripheral to the AI story were not necessarily peripheral at all; the map was missing part of their work.

A market trader in Cape Town, a baker in Lima, or a Sri Lankan plumber running a tiny business may appear only weakly exposed through a standard occupational lens, but look inside the way their work is actually organised and the picture changes.

Perhaps our maps were simply not built to see where that transformation could happen. If AI can supply some of the specialised capabilities that thin markets and small organisations do not, its transformative reach may extend much further across the occupational spectrum than we have assumed.

For economies that have struggled for decades to close productivity gaps, that is not only a story about jobs; it could become a new channel for much-needed catch-up growth.