Feedzai: Deep Technology Built in Portugal, Sold to Global Banks

Feedzai is the closest thing Portugal has to a deep-technology company built on domestic scientific talent. It was founded by data scientists and aerospace engineers, it competes on model performance against far larger American rivals, and it sells to the most demanding buyers in the world — the risk departments of global banks. This case study explains why fraud detection suits that profile and where the business is exposed. It is part of the Portugal Company Stories hub.

Why is fraud detection a genuinely hard technical problem?

Because it combines extreme scale, an adaptive adversary and asymmetric error costs. Systems must score billions of transactions with a latency budget measured in milliseconds, because a payment cannot wait for analysis. That alone rules out most machine learning architectures.

The adversary is the harder part. Fraud patterns change continuously, because the people committing fraud observe which methods are blocked and adapt. Static rule sets decay from the moment they are deployed, which is why the field moved to machine learning earlier than most areas of enterprise software.

Error costs are asymmetric and both directions hurt. Missing fraud costs money directly; blocking a legitimate transaction costs a customer relationship and generates support cost. Optimising that trade-off, rather than maximising detection alone, is what the product actually does.

How does a Portuguese company sell to global banks?

By being demonstrably better on measurable criteria, because banks buy risk software through rigorous evaluation rather than through relationships alone. A fraud system’s performance can be tested against historical data, and a vendor that detects more fraud at fewer false positives wins on evidence.

That is unusually favourable for a technically strong company from a small country. In categories where purchasing is driven by brand, incumbency or golf, a Portuguese challenger struggles. In categories where the buyer runs a bake-off with their own data, engineering quality is decisive.

The founders’ backgrounds matter here. A company started by data scientists and aerospace engineers, with academic roots in Coimbra, was positioned to compete on model performance from the outset rather than on sales coverage.

What did the 2025 funding round signal?

Domestic capital arriving at scale. The round of roughly $75m brought in new institutional investors including Portuguese firms Lince Capital, Iberis Capital and Explorer Investments, joining existing backers such as Oxy Capital and Buenavista Equity Partners.

That composition is notable. Portuguese growth-stage technology companies have historically been funded almost entirely by American and northern European venture capital, with domestic institutions absent. A round led substantially by Portuguese investors suggests the local capital market is developing capacity it previously lacked.

The valuation, reported above $2bn, also indicates the company grew through a difficult period for technology funding rather than raising a flat or down round, which many peers were forced to accept between 2022 and 2024.

Who are the competitors?

Large American software companies, in-house bank systems and specialised challengers. The category includes established vendors with substantial installed bases, several well-funded startups, and the internal fraud teams of the largest banks who build rather than buy.

The build-versus-buy dynamic is the structural competitive question. The largest global banks have the data science capability to build their own systems, and many do for core use cases while buying specialised capability at the edges. A vendor’s addressable market is therefore smaller than the total number of large financial institutions.

Regulation adds a second layer. Anti-money-laundering obligations are prescriptive and vary by jurisdiction, which favours vendors who can demonstrate compliance across multiple regulatory regimes — a barrier to entry that helps established players and hinders new ones.

How does generative AI change this business?

In both directions, which is the interesting part. It improves detection capability, enabling systems to reason about transaction context, generate investigation summaries for analysts and reduce the manual review burden that dominates operational cost in financial crime teams.

It also arms the adversary. Generative models make social engineering, synthetic identity creation and impersonation dramatically cheaper and more convincing, which increases fraud volume and shifts the attack surface from card testing toward authorised push payment fraud where the customer is manipulated into authorising the transfer.

The net effect is a larger market with harder problems. That favours specialists with deep model capability over generalist software vendors, which is the position a company built by data scientists occupies naturally.

What does Feedzai represent for Portugal?

Proof that deep technology can be built domestically. Unlike companies whose principal asset is a business model or a marketplace, Feedzai’s value rests on research capability, and that capability was developed in Portugal with roots in Portuguese universities.

It also demonstrates the value of the academic-industrial link that Portuguese policy has struggled to build systematically. The company emerged from a research environment, and its technical differentiation has remained its commercial argument — a pattern closer to the Israeli or Estonian model than to the marketplace-driven successes elsewhere in Portuguese tech.

Whether that is replicable is the open question. Portugal produces strong engineers and mathematicians; it has produced far fewer companies that commercialise research directly, and understanding why is more useful than celebrating the exceptions.

How does bank procurement actually work in this category?

Slowly and rigorously. A financial institution buying fraud technology runs a structured evaluation: technical requirements, a proof of concept against historical data, security and vendor risk assessment, regulatory review, integration planning and commercial negotiation. Twelve to eighteen months from first contact to contract is normal.

That process favours vendors who can demonstrate deployments at comparable institutions, because reference customers reduce perceived risk more than any feature comparison. Winning the first tier-one bank is disproportionately difficult and disproportionately valuable.

For a growth company it also means capital intensity. Supporting long evaluations requires solution engineers, data scientists and compliance staff working on opportunities that may not close, which is why fraud technology companies raise more capital relative to revenue than typical enterprise software firms.

What does the Portuguese research base contribute?

Access to a specific kind of talent at a specific cost. Portuguese universities produce strong graduates in computer science, mathematics, physics and engineering, and the country has research groups in machine learning and data systems with international standing.

For a company competing on model performance, that talent pool at Portuguese salary levels is a genuine structural advantage against American competitors paying Bay Area compensation for equivalent capability. It allows a larger research team for the same budget.

The limitation is seniority. Deep expertise in productionising machine learning at scale is scarcer in Portugal than in the United States, so companies typically hire senior leadership internationally while building depth locally — the same split that OutSystems applied in sales.

Why does anti-money laundering matter commercially?

Because the penalties are enormous and the obligations are non-negotiable. Financial institutions face substantial fines and supervisory action for compliance failures, which makes anti-money-laundering technology a purchase driven by risk avoidance rather than by return on investment.

That is a favourable dynamic for vendors. Budgets for regulatory compliance are more resilient in a downturn than budgets for growth initiatives, because the obligation does not disappear when profits fall.

The operational problem the technology addresses is false positives. Traditional rule-based monitoring generates enormous volumes of alerts, the overwhelming majority of which are not suspicious, and banks employ large teams to review them. Reducing that alert volume while maintaining detection is where machine learning delivers measurable savings.

How big is the financial crime technology market?

Large and growing faster than most enterprise software categories, driven by three forces: rising payment volumes as commerce digitises, increasing fraud sophistication, and regulatory obligations that expand rather than contract.

Real-time payment systems have accelerated it further. When settlement is instant and irreversible, fraud must be detected before authorisation rather than reconciled afterwards, which raises the technical requirement and the value of a system that can decide in milliseconds.

The result is a category where budgets grow through downturns, buyers are concentrated among large institutions and technical differentiation is measurable. That combination is unusual in enterprise software and explains why the category has attracted so much venture funding relative to its size.

What is the outlook for the company?

Growth in a structurally expanding market, with the usual constraints of an enterprise vendor selling to concentrated buyers. The 2025 round strengthened the balance sheet and brought domestic institutional investors onto the register, which supports continued investment in research and international expansion.

The strategic question is scale relative to competitors. Fraud technology rewards data volume and model quality, both of which improve with more customers, so market share compounds. A company at this size must keep winning tier-one institutions to stay in that compounding loop.

An eventual exit is likely to be acquisition by a larger payments or software group rather than a listing, which is the standard outcome for European enterprise software at this scale — and the same pattern described across the Portugal hub for companies that outgrow the domestic capital market.

Frequently Asked Questions