If for decades investors evaluated startups based on factors such as market size, growth speed and team quality, now a new variable is gaining strength: how protected a business is from being swallowed by artificial intelligence. Close to completing 20 years in the market, the Brazilian venture capital manager Trundle decided to develop a specific framework for this – a kind of resilience score.
The idea is to allow managers to objectively assess the chances of a startup being replaced not only by other AI native or AI first companies, but by the language models of giants such as OpenAI e Anthropicin addition to LLMs from big techs like Gemini (Google) and Copilot (Microsoft).
“The risk of disruption by another company has always existed. The difference now is the risk of disruption by a platform”, explains Daniel Chalfon, general partner at Trundle. “There are LLMs that will never solve 100% of the problem, but sometimes they solve 70%, 80%, with less cost. For some tasks, maybe it’s good enough.”
Study abroad
Upgrade your career!
Founded in 2008, the Trundle is in the final stretch of Journey V, its fifth fund, with eight investments made so far. The expectation is to close three or four more contributions before closing the cycle. Journey VI is scheduled for the second half of 2026 or early 2027. The manager’s focus is on startups in early stages, with Pre-Seed, Seed and Series A rounds.
This is not the first time that the house has created its own methodology: in 2024, the Trundle presented V multiple, a formula created to price early-stage startups based on gross margin.
For Chalfon, although the new AI risk framework does not directly affect V multiple, there is an impact of AI on companies’ gross margin that must begin to be considered.
“AI can affect gross margin, as it has been affecting the cost of tokens, for example. Furthermore, AI brings some transaction costs that didn’t exist before. Spotify, for example, is launching features with AI. This has a cost. Will they charge for it? Gemini’s photo editing is done in the cloud, but Google doesn’t charge for it. So AI affects V multiple in the sense that it accelerates revenue, but also reduces gross margin”, ponders the investor.
The seven criteria
The framework classifies companies into three bands – most protected (Score 2), medium risk (Score 1) and at risk (Score 0) – based on seven criteria that requalify the so-called moatsthe classic competitive advantages of the venture capital market, from the perspective of the new technological era.
Data Advantage is the first pillar. Generalist AI models have already been trained with almost everything on the internet. This means that any product based on public or easily purchasable data is exposed. The real protection lies in the so-called “walled gardens”: proprietary data, regulated or generated by the operation itself, that no model can replicate. The more exclusive and difficult to copy, the higher the barrier.
Lock-in institucional evaluates the friction generated by changing the solution – whether it creates regulatory, legal or compliance risk for the customer. “You’re unlikely to change software that runs the bank because Claude arrived. But maybe you’ll change a simpler CRM”, explains the partner.
Continues after advertising
Embedded in transaction or workflow measures how inseparable the product is from the customer’s critical operation. When software is the revenue-generating transaction rail, shutting it down stops billing immediately.
Distribution control analyzes whether the company controls access to the buyer in a way that is difficult to disintermediate. “Sales is sales, especially in B2B. Having a differentiated channel, a relationship that is difficult to change, is a favorable point”, says Chalfon. “Increasingly, distribution via proprietary channels becomes a moat.”
Local complexity considers the particularities of the Brazilian market, such as specific legislation, bureaucracy and constantly changing tax rules. “Brazil is a large market, and local specificities are important. Generalist software has difficulty here.”
Continues after advertising
Network effect differentiates user base growth from the true network effect. “If you have a tool that gets better as it grows, the effect perceived by the customer guarantees lock-in”, he explains. SLMs (small language models) are cited as a case in which this advantage would be realized more quickly.
Finally, adaptation speed it is the dynamic criterion of the matrix, and the only one capable of changing the classification range. It assesses leadership’s ability to not only recognize disruption, but actively incorporate AI into the operation.
“The main layer of change depends on leadership. That is, if leadership recognizes the need to change and test new ways to incorporate AI, whether it reacts slowly. We look at the productivity of personnel, the department, and when an entire process changes because of AI”, notes Chalfon.
Continues after advertising
Content produced by
