By Cesar Augusto Taborda Lima
Whoever buys a tech company in Brazil is buying, above all, a set of rights over code, data and models, and due diligence exists to confirm that those rights sit where the seller says they do. The problems that surface are nearly always the same. There is code written by a founder before the company existed, or by collaborators with no contract defining what they were doing. There are databases built without an adequate legal basis, models trained on data that could not be used that way and open-source or AI-vendor licences incompatible with the product. What the investigation finds becomes price, a condition to closing or an indemnity clause. And there are obligations outside the contract, such as the CADE merger filing and the declaration of the investment to the Central Bank, that do not wait for signing to exact their price.
Art. 4. Unless otherwise agreed, the rights to a computer program developed and produced during a contract or statutory relationship expressly intended for research and development, or in which the activity of the employee, service contractor or civil servant is provided for, or which arises from the very nature of the duties of those relationships, shall belong exclusively to the employer, the contracting party or the public body.
Law 9,609/1998 (Brazil’s Software Law), art. 4, free translation.
Who owns the code
Brazil’s Software Law is generous to the company. A program written by an employee, a service contractor or an intern belongs to whoever hired them, provided the contract has development as its object or the activity follows from the nature of the role. The trouble lies at the edges of that rule. Code a founder wrote in the months before the company was incorporated, when there was no contract with anyone, belongs to the founder personally until a written assignment transfers it to the company. A program made with no connection to the contract and without the company’s resources belongs to whoever made it (art. 4, § 2), and a developer who worked for months without a written contract leaves the buyer the task of proving what the engagement covered.
Everything that is not software follows another statute. Interfaces, illustrations, copy, videos and visual identity are works protected by the Copyright Law, Law 9,610/1998, which has no work-for-hire rule: transferring the economic rights requires a written contract (arts. 49, II, and 50). A trademark registered with the INPI in a partner’s name and a domain held in a personal account are routine findings, easily fixed by a transfer before closing, provided someone looks for them.
How the team was hired also weighs on price. Brazilian startups often engage much of their engineering staff as contractors through their own companies, and the Supreme Court has yet to decide, in Theme 1,389, whether labour courts may rule on fraud in those contracts and who bears the burden of proof; since June 2026, cases have been held at the regional courts awaiting the thesis. Until it comes, the contingency exists and has to be estimated, because in a share deal it stays inside the company.
Open source and third-party models
Almost every tech product is built on open-source libraries, and most of them cause no trouble. Some licences, however, require derivative code to be distributed on the same terms, and some extend that requirement to software offered over a network. One such library at the core of a closed product can force open the code the buyer thought it was acquiring exclusively. An automated scan of the repository settles much of the question within days.
With artificial intelligence, the question changes in kind. Many startups sell as their “own model” a layer of prompts and integration on top of a large provider’s model, reached through an API. That changes what is being bought: the central asset is licensed, the terms can change at any time, large providers commonly forbid using outputs to develop competing models, and the cost of each call goes into the margin. Where the model really is proprietary, the investigation returns to the layers I discussed in the article on AI development contracts: who owns the weights, what data they were trained on and under what licence the base model sits.
The data
In a share deal, the company carries its entire data protection liability with it. Brazil’s General Data Protection Law (LGPD) provides for fines of up to 2% of revenue in Brazil, capped at R$50 million per infringement (art. 52, II), alongside civil liability to data subjects (art. 42). The investigation looks for the legal basis of each relevant dataset, the record of processing activities, the incidents that occurred and what was done about them.
Model training deserves a chapter of its own. In July 2024 Brazil’s data protection authority, the ANPD, suspended by precautionary measure the use of Instagram, Facebook and Messenger users’ data to train Meta’s AI, under a daily fine of R$50,000, for lack of transparency, the processing of children’s and adolescents’ data and the obstacles placed in front of those who wished to object (Decision 20/2024). The suspension was lifted in August, after the company agreed to exclude under-18s, de-identify the data and offer each user a simple opt-out. For a large platform it was an adjustment. For a startup whose model is its main asset, an order like that could mean retraining the product from scratch, and that risk has a price.
Then there is the day after closing. A foreign buyer that integrates the group’s systems, servers and databases is making an international transfer of personal data, and since 23 August 2025 that transfer must be covered by the standard contractual clauses of ANPD Resolution 19/2024, whose text must be kept in essence, or by another accepted mechanism.
Outside the contract: CADE and the Central Bank
The transaction must be notified to CADE, Brazil’s competition authority, before closing when one of the groups involved had gross revenue in Brazil of at least R$750 million in the previous year and the other at least R$75 million (Law 12,529/2011, art. 88). Closing before approval leads to nullity and a fine of R$60,000 to R$60 million (§ 3). Most startup acquisitions fall below those figures, but CADE may require a non-notifiable transaction to be submitted within one year (§ 7), a power that matters precisely when a large tech company buys a small competitor. A public consultation opened in September 2026 proposes counting, on the seller’s side, only the revenue of the company being sold (Cescon Barrieu); for now, it is a proposal.
A foreign buyer’s investment must be declared to the Central Bank by the Brazilian company that receives it, in the SCE-IED system, whenever there is a transfer or movement of US$100,000 or more, besides periodic declarations by asset size (Law 14,286/2021 and Central Bank Resolution 278/2022, per Machado Meyer). It is the record that supports, years later, dividend remittances and the return of capital.
From finding to contract
Structure comes first. Buying assets instead of shares looks like protection from liabilities, but the buyer of a going concern answers for prior debts duly recorded in the books (Civil Code, art. 1,146), and tax and labour succession rules reach that purchase (National Tax Code, art. 133; Labour Code, arts. 10 and 448). In tech, share deals tend to prevail, because contracts, licences and staff stay where they are, and the liabilities are then handled by the contract.
Each finding has a destination. What can be fixed before closing becomes a condition: assignment of the founder’s code, transfer of the trademark, signature of the standard clauses. What can be quantified goes into the price. What is uncertain goes into specific representations and warranties on ownership, data, open source and model training, with a dedicated indemnity and part of the price held in escrow for as long as the risk may surface. Between businesses, the Civil Code directs courts to respect that allocation of risk (art. 421-A). The earn-out, common when buyer and seller disagree on value, calls for special care in tech: after closing, the buyer usually integrates the product and changes the metrics, and that is exactly where the dispute begins. The contract must say how each metric is measured and what the buyer may and may not do with the business during the period. I covered governing law, forum and arbitration in these transactions in the article on contracts with foreign companies.
It seems to me that due diligence on a tech company is decided by a question asked early: what, exactly, is the buyer paying to take away. If it is the code, the investigation starts with the chain of title. If it is the data, with the legal basis of each dataset. If it is the model, with what went into training and the licences it depends on. That question organises the work and spares the parties a long report that changes not a single line of the contract.