the model we built with forgone taxes and what comes next

to understand what is happening in the uruguayan tech sector you have to start where people rarely start: with the money the state decided not to collect.

since 1998, with successive expansions through 2023, uruguay built one of the most generous tax regimes on the continent for the software industry. companies that develop their own software or software for third parties are exempt from irae, the corporate income tax whose nominal rate is 25%. the sector's effective rate can be zero. the dividends those companies distribute pay no irpf. companies installed in tech free trade zones, like zonamerica or aguada park, have full exemption from all national taxes. service exports pay no vat. processing equipment pays no wealth tax. and since 2023 there is a special regime to attract technical talent from abroad that reduces personal social security and tax burdens for five years.

uruguay's total tax expenditure is around 6.6% of gdp, some 5.3 billion dollars a year that the state stops collecting. it is the highest in latin america and double the regional average.

the tech sector is not the only beneficiary. agriculture has reduced rates and simplified regimes that make it effectively pay much less than other industries. pulp operates in free trade zones: upm invested 3.4 billion dollars in its second plant and pays an annual fee of 7 million. in uruguay the exemption is not the exception, it is the central mechanism of industrial policy. what changes between sectors is who negotiated better and when.

but there is a structural difference between tech and agriculture or pulp. agriculture exists because there is land. pulp exists because there are forests and a river. tech exists because there is human capital, and that human capital is not tied to the territory the way a cow or a tree is. that makes the incentives, in theory, more necessary to retain it. it is exactly the argument the sector repeats every time someone suggests reviewing the regime.

the problem is what was built with those incentives.

for twenty years, the exemptions fundamentally financed a nearshoring model: companies that hire uruguayans to develop software that gets used somewhere else, for clients who decide from somewhere else, with intellectual property that stays somewhere else. cuti acknowledges it without euphemisms: 70% of the activity and 80% of exports correspond to software factories. the state exempted taxes and received quality employment and exports in exchange, but no intellectual capital accumulated in uruguay. the knowledge left with the projects.

the model worked well while it worked. between 2000 and 2024 the sector's revenue grew from 200 million dollars to 3.681 billion. 20,500 direct jobs were created with average salaries of 2,300 dollars a month. but software factories have a fragility that aggregate numbers hide: decisions are made elsewhere. sabre laid off 200 people in 2023 and another 150 in 2025, communicated with a video from dallas. there was no deterioration of local results, there was no performance review. there was a corporate decision in some efficiency committee in the united states and the montevideo team found out along with the teams in bangalore and warsaw. 81% of the sector's revenue is concentrated in 22 companies, most of them subsidiaries or contractors of multinationals. that is not a tech industry. it is a highly specialized services sector with a structural dependence on external demand.

the irruption of artificial intelligence did not change this logic, it accelerated it.

the work that software factories bought in uruguay was, for the most part, describable work. development cycles, testing, integration, support, maintenance. work that follows a process, that has defined inputs and outputs, that can be documented. exactly the kind of work that language models and coding agents are beginning to replace or to compress into much smaller teams. a company that used to need 40 developers for a project of medium complexity can do it today with 12 if it uses the available tools well. that explains the layoffs the sector has been processing since 2023, and there is no reason to assume the process stops.

the profile of the resource the market is looking for now is different from the one formed under that model. knowing how to code is no longer enough. what is demanded is the ability to work at the interface between business and technology: understanding what problem is being solved, designing how an agent can solve it, validating that it works, iterating. it is a combination of product judgment, technical ability and process understanding that does not yet have a clear name in resumes or in job postings, but that the companies who hire well already know they are looking for.

the problem is that there is no institution in uruguay that trains that systematically. the university offering is structured to produce the engineer the previous model needed. fing trains good software engineers in the classic sense. ort has degrees aligned with the corporate market. there are bootcamps that teach frameworks and languages. but none of these formats is designed to produce someone who knows how to orchestrate agents, think in terms of end-to-end automation, work with ai tools as an execution layer instead of as an object of study. a university's response time to modify its curriculum, validate it and implement it is between four and seven years. by the time there are graduates trained for this context, the context will have already changed again.

what remains is a gap that nobody closes institutionally. some individuals solve it on their own. some companies retrain profiles internally. but there is no program, there is no policy, there is no actor whose explicit objective is to train at scale the resource the sector is going to need in the next five years.

in that context, it is worth observing how the other side of the equation responds: the companies that acquire technological capabilities instead of transforming internally.

white pearl technology group ab is a swedish it services holding listed on nasdaq first north, with around 50 million dollars in annual revenue and 850 employees distributed across 37 subsidiaries in 30 countries. in 2024 it acquired 50% of ataraxy (our company). the transaction was announced with a press release describing the acquisition as part of its expansion strategy in ai and latam. with great fanfare.

wptg's numbers look good on the surface. revenue grew 65% year over year, ebitda margin reached 16.9%, and the stock rose 124% in a year. those are the numbers a retail investor in stockholm looks at and concludes that something is working. the problem is what those numbers do not measure.

according to carlsquare's analysis, 70% of wptg's revenue growth in 2025 came from acquisitions, not from organic growth. the company completed at least seven transactions in two years: a 15-person sap consultancy in sweden, a 100-person it staffing firm, a digital marketing agency, an indian voice ai startup for 300,000 dollars, and ataraxy. the pattern is not transformation, it is consolidation. small profitable companies are bought, their revenue is added to the consolidated figures, and the result is presented as the growth of a diversified technology group.

the narrative that accompanies all of this is aggressive. wptg talks about "revolutionizing the it services industry in the next two or three years", about a technology called "digital agent model" that "will fundamentally change how consulting firms deliver services", about something named "true multiverse" that integrates cloud, legacy platforms and frontier technology. these are the press releases that steve blank, a stanford professor, described in harvard business review as the defining trait of innovation theater: "big headlines with few tangible details". the company has also reissued its 2023 annual report over inconsistencies and corrected its 2025 quarterly report over accounting errors, which for a company of this size is unusual.

this is not a specific critique of wptg. it is the description of a pattern documented with decades of evidence. the circus cannot stand in for innovation.

eds, which literally invented it outsourcing, reported 21.5 billion dollars in revenue and 1.36 billion in net profit in 2001. hp bought it in 2008 for 13.9 billion. four years later it took an 8 billion writedown, acknowledging that it had destroyed value. the fixed-contract model for managing on-premises data centers was exactly what the market was beginning to eliminate with aws.

dxc technology was born from the merger of csc and hpe enterprise services in 2017, with 25 billion in revenue and 170,000 employees, promising more than 1 billion in synergies. its stock started at 74 dollars, reached 92 in 2018, and today trades around 13. revenue fell 48%. it was removed from the s&p 500 in 2023. three ceos in seven years. zacks' analysis rates it a "strong sell". the conclusion it left engraved in the sector: two declining businesses do not create a growing one.

atos, the european it services giant, fell from a valuation of 15 billion dollars to fractions of that value before requiring emergency injections from the french government to avoid disappearing.

the pattern is the same in every case. profitable companies, with good quarterly numbers, with digital transformation narratives, that did not make the real operational change while the results still allowed it. when the change became urgent, they no longer have the agility to make it.

academic research has a name for this phenomenon. levinthal and march formalized it in 1993 as the success trap: organizations that perform well for extended periods develop such a strong dependence on their current practices that they become structurally incapable of exploring alternatives. not because they do not see the change, but because their entire management infrastructure, their incentive systems, their career paths and their organizational values are built to optimize what already works. dorothy leonard-barton called it "core rigidities": the very capabilities that made a company successful become the chains that immobilize it when the environment changes.

in it services, this translates with precision: a company whose core competence is managing delivery teams based on labor arbitrage cannot transform into an agent-first organization without first destroying the measurement systems, incentives and culture that made it profitable. and as long as it remains profitable with the old model, it has no incentives to destroy them.

at some point in 2023 that logic stopped being abstract and became operational. for sidetool and ataraxy it became clear that continuing to do the same thing was a slow way of shutting down. it was not an epiphany. it was an accumulation of signals all pointing in the same direction: clients were asking for more with less, cycles were compressing, the tools did in hours what used to take weeks. the decision to reformulate the model was not strategic in the elegant sense of the term. it was a survival decision.

today we operate under a logic we call agent-first. the starting point of any project is not how many people we need but how much an agent can do before a person intervenes. where eight people used to be needed to cover a full cycle, today we run with three or four: someone who understands the business problem in depth, someone who designs and orchestrates the agents, someone who validates that the output is reliable. the rest of the cycle is executed by agents, not as assistants but as a production layer.

the clients we work with are almost all international, mostly in north america and europe. not because we decided to ignore the local market, but because they are the ones with the urgency and the context to understand what is being offered. a company that is evaluating how to automate a collections operation, or how to process clinical documentation at scale, or how to reduce cycle time in procurement, immediately understands the value of a small team that delivers in weeks and takes responsibility for it working, not for how many hours it billed.

the software factory model does not compete in that space. not because it is inferior in technical quality, but because it is structured for a different question. the software factory answers: i need development capacity, how much does a qualified person cost in montevideo versus warsaw versus bangalore? that is a question ai is also answering, and it answers it cheaper than any human team.

reconverting while you keep operating has a cost that the tech pivot narrative usually omits. it involves decisions about which projects to accept and which not to, about which profiles no longer have a place in the model, about how to communicate inward and outward that the organization is a different one without generating more noise than necessary. it is not reinventing yourself from scratch. it is changing the engine while the car is moving.

uruguay bet for twenty years that exemptions were enough to build a tech sector. it did reasonably well with that model. but the model it financed is the one under pressure today. the transition toward what comes next does not have the institutional actors it should have, and the organizations making the move are doing it alone, without a net, solving in real time a question that the educational system and public policy have not yet finished formulating.

the good results of past quarters are not evidence that the model works. they are evidence that the moment when it stops working has not arrived yet. that distinction, in this context, is everything.