A startup that says it cut violent crime by more than a quarter in six months would normally be met with raised eyebrows. A startup that says it did so while raising one of the largest seed rounds in Latin American history, from the investors behind Anthropic and Uber, demands a closer look. Pax just made both claims at once, and the implications reach far beyond São Paulo. The same platform that solves homicides is also the most complete surveillance infrastructure most of these cities have ever installed, and the round just poured fuel on that fact.
What Actually Happened
Pax, a São Paulo-based AI company building real-time intelligence software for public safety, raised a $40 million seed round co-led by Greenoaks and Benchmark, one of the largest seed investments in the history of Latin America. The company pairs that funding announcement with an operational claim that is far more striking than the dollar figure: in its first large-scale deployment, Pax says it cut violent crimes by 27 percent, doubled police efficiency, and lifted public perception of safety by 59 percent, all within six months. For a seed-stage company, leading with outcomes rather than a roadmap is unusual, and it is the entire pitch.
The product is a data-fusion platform. Pax connects cameras, police records, and criminal databases into a single intelligence system that turns fragmented, siloed information into investigative output. It reviews video footage, generates leads for detectives, and answers questions through a conversational interface, so an officer can interrogate a sprawling evidence base in plain language rather than filing requests across incompatible systems. The pitch is that most police forces are not short on data. They are drowning in it, with footage and records scattered across agencies that do not talk to each other, and Pax positions itself as the connective layer that makes the existing data usable.
The traction numbers are concrete enough to check over time. Pax says police forces using its platform have resolved more than 2,000 criminal cases, including homicides, armed robberies, and vehicle thefts, across more than 30 cities. The investor roster underlines how seriously the capital markets are taking it. Greenoaks has backed Coupang, Brex, Revolut, Flock Safety, and Anthropic, while Benchmark's portfolio includes Uber, eBay, Instagram, and Snap. A seed round co-led by two firms of that caliber, at this size, signals conviction that Pax is building a category rather than a feature.
Why This Matters More Than People Think
Latin America carries a disproportionate share of the world's violence. The region is home to a small fraction of the global population but a large share of its homicides, and the cost of that violence, in lives, in lost investment, in emigration, is staggering. A tool that credibly bends the violent-crime curve in a major Brazilian deployment is addressing one of the hardest and most expensive problems any government faces. If the 27 percent figure survives independent scrutiny across more cities, Pax is not selling software efficiency. It is selling a measurable reduction in human suffering, which is a different and far larger market.
The structural insight is that public safety has been starved of modern software. Police departments worldwide run on legacy records systems, disconnected camera networks, and manual evidence review that consumes detective hours by the thousand. Applying frontier AI to that mess, connecting the data and letting officers query it conversationally, is the kind of unglamorous, high-leverage automation that transforms a workflow without inventing a new science. The same playbook that turned messy enterprise data into searchable intelligence for corporations is now being pointed at the state, and the state spends more and moves slower, which means a longer runway for whoever gets it right.
There is a market-timing point that explains the round size. Conversational AI good enough to interrogate video and records at scale did not exist three years ago, and the agencies that buy this software move on multi-year procurement cycles. A company that lands early, proves outcomes, and embeds into a city's operations becomes very hard to displace, because switching costs in government are punishing and incumbency compounds. Greenoaks and Benchmark are not paying $40 million at seed for current revenue. They are paying for the chance to own the default public-safety intelligence layer across an entire region before anyone else arrives.
The economic stakes for governments make the budget math favorable in a way that is easy to miss. Violent crime imposes costs that dwarf any software contract: emergency response, courts, incarceration, lost productivity, capital flight, and the tax that insecurity places on every business decision in a region. Economists routinely estimate the cost of violence in Latin America at several percent of GDP. Against that backdrop, a municipal contract for an intelligence platform is a rounding error, which means Pax is selling into a budget line where the return on investment, if the crime reduction is real, is overwhelming. That asymmetry is what lets a public-safety software company command venture-scale economics rather than the thin margins usually associated with selling to the public sector.
The Competitive Landscape
The most obvious comparison is Flock Safety, the US license-plate-recognition and camera-network company that Greenoaks also backs, valued in the billions. Flock proved that police agencies will buy AI-driven surveillance infrastructure and pay recurring fees for it. Pax extends the model into a harder, more fragmented market and broadens the scope from cameras alone to a unified intelligence layer across records and footage. Greenoaks backing both is not a coincidence: it is a thesis that the public-safety software category is large, durable, and underbuilt, and that the winners will be vertical-specific rather than generic.
Palantir is the looming shadow over this entire space. Palantir built a multibillion-dollar business fusing data for government and defense, and it demonstrated that the unsexy work of connecting institutional databases is worth more than most consumer AI. Pax is, in one reading, a focused, AI-native, Latin-America-first answer to the question Palantir raised: what if the data-fusion layer for the state were built from scratch around modern language models and aimed squarely at street-level policing rather than intelligence agencies? The regional focus is the wedge, because Palantir and the US incumbents are not optimized for Brazilian agencies, budgets, or languages.
The historical parallel is the spread of CompStat in 1990s New York. CompStat was not an algorithm. It was a system for connecting crime data to accountability, and its proponents credited it with part of the era's crime decline while critics argued it encouraged metric-gaming and aggressive enforcement. Pax is CompStat with a language model: the same promise of data-driven policing, the same potential for dramatic reported results, and the same unresolved questions about what gets optimized and who bears the cost. The technology is new, but the governance debate it reopens is three decades old.
Hidden Insight: The Surveillance Bargain Nobody Voted On
The uncomfortable truth beneath the 27 percent is that crime reduction and surveillance expansion are being sold as a single package, and societies rarely get to price them separately. A system that fuses every camera, record, and database into a queryable intelligence layer is precisely the infrastructure of a surveillance state when pointed in a different direction. The same capability that helps a detective solve a homicide can, with a change of operator or intent, track dissidents, profile neighborhoods, or monitor a population. Pax is being deployed in democracies, but the tool does not know what kind of government holds it, and infrastructure outlasts the administration that installs it.
The bear case, however, is not only about civil liberties. It is about whether the headline numbers mean what they appear to mean. Critics argue that a 27 percent drop in violent crime over six months in selected cities is exactly the kind of result that can reflect deployment in already-improving areas, changes in reporting rather than reality, or short-term effects that regress once the novelty fades. Crime statistics are notoriously sensitive to how they are counted, and the incentive to show a flagship customer a dramatic number is enormous at the seed stage. The risk the market may be underpricing is that the causal claim, Pax reduced crime, is far harder to prove than the correlational one, and that independent evaluation will soften the figure.
There is also the bias problem that has haunted every predictive-policing system before this one. Models trained on historical police data inherit the patterns of historical policing, which in most countries means over-enforcement in poorer and minority neighborhoods. If Pax generates leads and prioritizes attention based on data shaped by past bias, it can launder that bias into something that looks objective because a machine produced it. Skeptics point out that the most dangerous surveillance tools are the ones that work well enough to be trusted, because trust suppresses the scrutiny that would catch the failures. A tool that visibly reduces crime is far harder to regulate than one that obviously does not.
The deepest strategic insight is that Pax is building the operating system for state power in a region where institutions are still being shaped. Whoever provides the intelligence layer that police, and eventually other agencies, depend on accrues quiet, structural influence over how the state sees its citizens. That is a position of extraordinary leverage and extraordinary responsibility, and it is being established now, fast, with venture capital urgency and before any regulatory framework exists to govern it. The window in which society could deliberately decide the rules for this kind of infrastructure is open today and closing quickly, and the funding round just accelerated the clock.
There is a venture-capital reading of the round that sharpens the stakes further. Greenoaks and Benchmark did not co-lead a $40 million seed for a product that merely sells well. They priced it as a potential category monopoly, the public-safety equivalent of what Palantir became for defense, with the regional moat and switching costs to match. That ambition means Pax will be under pressure to expand fast, win flagship cities, and entrench before regulation or competition catches up. Speed is rational for the company and dangerous for the public, because the faster a surveillance layer embeds into the machinery of the state, the less opportunity there is for deliberate, democratic debate about whether and how it should exist. The incentives of the cap table and the interests of the citizenry are not aligned here, and that gap is where the real story will play out.
What to Watch Next
In the next 30 days, watch for independent verification of the crime numbers. The claims are specific and checkable, so the question is whether academic criminologists, journalists, or the cities themselves publish data that confirms or qualifies the 27 percent. A peer-reviewed or third-party evaluation would transform Pax from a compelling pitch into a proven category leader. Silence or pushback from the deploying agencies would be a yellow flag that the marketing outran the evidence.
Over 90 days, watch the expansion map and the contract structure. The tell is whether Pax converts pilots into multi-year municipal or state contracts beyond its initial deployments, and whether it expands to other Latin American countries with different legal regimes around surveillance. New flagship cities, especially in Mexico, Colombia, or Argentina, would confirm a regional land grab. Watch also for the first organized civil-liberties response, because the arrival of serious privacy and human-rights scrutiny is the marker that the category has become consequential enough to fight over.
By 180 days, the decisive question is governance. Watch whether Brazil or any deploying jurisdiction moves to regulate AI-driven police surveillance, and whether Pax positions itself ahead of that wave with transparency and audit commitments or resists it. The companies that survive in government AI long-term are the ones that build oversight into the product before the law demands it. Whether Pax treats civil-liberties guardrails as a feature or a friction will determine if it becomes durable public infrastructure or a cautionary tale about moving fast in the most sensitive market there is. The companies that endure in this space will be the ones that prove their numbers, open their systems to audit, and accept that selling tools to the state carries obligations that selling software to businesses never did.
Pax is selling less crime and more surveillance in the same box, and societies almost never get to buy just one.
Key Takeaways
- $40M seed co-led by Greenoaks and Benchmark, among the largest seed rounds in Latin American history.
- 27 percent cut in violent crime, doubled police efficiency, and a 59 percent rise in perceived safety in six months, per Pax.
- 2,000+ cases resolved across 30+ cities, spanning homicides, armed robberies, and vehicle thefts.
- A Flock Safety and Palantir comparison: Greenoaks backs both Pax and Flock, betting public-safety AI is a large, durable category.
- The core tension: the same data-fusion layer that solves crimes is the infrastructure of a surveillance state in different hands.
Questions Worth Asking
- If a tool cuts crime by 27 percent and expands surveillance at the same time, how should a society decide whether the trade is worth it, and who actually gets to decide?
- When a machine generates police leads from historically biased data, does the objectivity it appears to add make the bias easier or harder to challenge?
- Who should own the intelligence layer that the state uses to see its citizens, and what happens to that infrastructure when the government holding it changes?
