Brazil Generative AI in BFSI: Pix and Open Finance Already Built the Substrate

The Brazil generative AI in BFSI market has no defensible size figure, and Brazil is still the best country in the world to build a banking or insurance AI product. Every regulated AI product needs two inputs a founder normally cannot buy. Payment data with near universal coverage, and customer financial data that moves on the customer's instruction instead of on a bilateral deal with a bank. Brazil shipped both as public infrastructure before the model layer got cheap.

Pix cleared 79.8 billion transactions worth R$ 35.36 trillion in 2025. Open Finance Brasil reported 208.79 million active consents on 31 July 2026. A founder in New York or Frankfurt assembles that substrate one integration at a time and pays for each one. In Brazil the customer hands it over by law, in a standard format, at no cost.

Avante Ventures is a venture studio building AI-native companies in Brazil and Latin America. What follows is our read on where the openings actually sit across banking, credit and insurance, what the moat is once you strip out the model, and which parts of this are hard enough to sink a team that walks past them.

The Brazil generative AI BFSI market, with dated numbers

Start with the weakest part of the case, because pretending otherwise is how people underwrite badly. No research house publishes a defensible Brazil-specific figure for generative AI in BFSI. What circulates is a global BFSI segment estimate multiplied by a country weighting, and estimates for the same country and the same base year diverge by more than 2x. A category whose definition is still moving looks exactly like that from the outside.

One number held up at source, and it covers the whole country rather than the vertical. IMARC Group sizes the Brazil generative AI market at USD 371.2 million in 2025, reaching USD 1,481.5 million by 2034 at a 16.63 percent CAGR. Useful as a directional signal. Useless as an underwriting input.

Observed spend is the stronger evidence, and it is public. FEBRABAN and Deloitte put the Brazilian bank technology budget at R$ 47.8 billion in 2025, up 13 percent from R$ 42.3 billion the year before, with investment in AI, analytics and big data growing an estimated 61 percent and Open Finance investment up 65 percent. More than 80 percent of surveyed banks already use generative AI, reporting an average 11.4 percent efficiency gain, and 38 percent report gains above 20 percent. The survey covers 20 banks representing 85 percent of Brazilian banking assets.

A forecast is one analyst's model of a future nobody has seen. A budget line is money already committed by the institutions holding most of the country's banking assets. When the two disagree, take the budget. It is the only one of the pair that has already survived a risk committee.

The macro backdrop is consistent with that. Services account for roughly 70% of Brazilian GDP on IBGE's value-added basis, while the World Bank series puts services value added at 59.7 percent of GDP in 2025. Both are correct. GDP additionally counts net taxes on products, which is why the value-added share reads higher. Financial services sit inside that block and digitize faster than the rest of it, a pattern we trace in more depth in our analysis of the Brazilian services economy opportunity.

Why Pix and Open Finance already built the data substrate

Permissioned financial data in Brazil is portable by regulation rather than by partnership. That one sentence is the thesis, and it is the part that is not true in the United States or most of Europe.

Pix, launched by Banco Central do Brasil in November 2020, handled 79.8 billion transactions worth R$ 35.36 trillion in 2025, a 33.6 percent increase in value over R$ 26.24 trillion in 2024. That averages roughly 219 million transactions per day, with peak days above 300 million. More than 170 million people use it, about 93 percent of the Brazilian adult population, and person-to-business volume now runs at 44 percent of the total.

For a builder that means merchant cash flow, consumer spending behaviour and counterparty graphs sit on a single national rail with near universal coverage. No card network fragmentation. No batch settlement lag. No bureau-only view of a thin-file borrower who has never held a credit card but takes 400 Pix payments a month.

Open Finance is the second leg and the more consequential one. It is mandatory for large institutions and customer-permissioned by design, not a voluntary API programme a bank can quietly deprioritise. The trajectory is the tell. FEBRABAN counted 62 million active consents in January 2025, 44 percent above the 43 million a year earlier. Eighteen months later the dashboard reads 208.79 million. That curve is a function of a mandate, not of anyone's product roadmap.

Drex is the third leg and it is not ready. It remains a restricted pilot involving banks, fintechs, B3 and Visa, with broad rollout targeted for 2027 after Banco Central reoriented the programme because the original distributed-ledger architecture had not met bank-secrecy and privacy requirements. Treat Drex as optionality. Anyone selling it as a 2026 dependency is selling a roadmap, not a product.

• 208.79 million active consents as of 31 July 2026, spanning transactional, credit, investment and insurance data.
• More than 2.3 billion successful API communications per week, with FEBRABAN-associated banks having invested over R$ 2 billion in the regime.
• Banco Central's 2025 to 2026 agenda extends the scope further, into salary and investment portability coordinated with the CVM and credit portability for unsecured lending.

The openings across banking, credit and insurance

The sector is not the opportunity. The decision is. Ranking matters more than listing here, because somebody with a larger balance sheet is already building three of the four openings below. The order follows how cleanly the Brazilian substrate maps onto each one.

Insurance earns its place on arithmetic. Brazilian insurance penetration was 3.3 percent of GDP in 2024 per MAPFRE Economics, while total insurance penetration across OECD countries averaged 6.2 percent in 2024 per the OECD. Both figures measure gross direct life and non-life premiums over GDP, so the comparison holds. Brazil sits at roughly half the developed-market level on a base that keeps expanding. CNseg projects R$ 808 billion in 2026 revenue, up 5.7 percent, and the industry's own development plan targets 10 percent of GDP by 2030. Low penetration on a growing base is a growth argument, not a maturity argument.

None of the first three openings below is a chat interface. Each one is a decision that has to leave a paper trail, which in Brazil is the only shape of AI product that reaches production at all.

• Credit underwriting where the model writes the reasoning a regulator can audit. Open Finance supplies permissioned cash-flow data on a thin-file borrower and Pix supplies merchant revenue seasonality. The product is not a score. It is a decision memo with the evidence trail attached, because that is what an examiner and a disputing customer both ask for.
• KYC and AML triage. High volume, high false-positive rate, and a human review queue that is the actual cost centre. The model drafts the narrative and assembles the evidence. The analyst keeps the decision.
• Insurance claims adjudication and underwriting. The strongest growth argument of the four, and the one we weight most heavily.
• Servicing. The lowest defensibility of the four and the one every incumbent is already building. Enter here only as a wedge into the other three.

Why permissioned data plus a compliance posture is the moat

Now the harder question. A chat layer over a core banking system is not a company. It is a feature the core provider ships next quarter, bundled, at zero incremental price.

The defensible asset has two parts and neither of them is the model. The first is the outcome data. Open Finance makes the raw feed legally portable, which lowers the barrier to entry for everyone including the competitor who copies the demo. What is not portable is what a live book produces. Every claim declined and later paid, every loan approved and later in default, every human override of the model, is a training label nobody can buy at any price.

The second part is the compliance posture treated as a product surface rather than as documentation. In BFSI the audit trail is the thing the buyer is purchasing. A copilot that emits a versioned, reproducible reasoning record for every decision, carrying model version, input data lineage, confidence and the reviewer's identity, sells into a regulated institution. One that emits an answer does not get past information security.

Local operating knowledge belongs in the same column as the data. Brazilian businesses spend an average of 1,501 hours a year meeting tax obligations, roughly five times the Latin America and Caribbean average and close to ten times the OECD high-income average. In BFSI that drag lands twice. Once in your own back office, and again inside the product, because tax treatment and documentary requirements are embedded in credit files, claims files and KYC packages. A foreign entrant prices the first cost and never sees the second one coming.

In Hamilton Helmer's terms the durable powers here are cornered resource and switching costs. The cornered resource is the outcome data on a live book. The switching cost is a decision log already embedded in the customer's own regulatory filings. Scale economies and network effects are weak in this category. Do not claim them in a deck a sophisticated investor will read.

How an underwriting copilot compounds toward a fund

This is the copilot to data to fund flywheel, applied to BFSI, and BFSI is the sharpest version of it because the outcome data carries a price. A loss curve on a specific book in a specific segment is not a nice-to-have analytics asset. It is the input a reinsurer or a credit facility underwrites against, which means the same dataset that improves the software also lowers the cost of capital for the vehicle it eventually feeds.

The capital market is currently paying for that order. Brazilian fintechs raised USD 2.77 billion across 106 rounds in 2025, against a comparable total spread over 244 rounds in 2021. Money did not leave Brazilian financial technology. It concentrated into fewer and larger cheques for teams that can show unit economics. A venture that arrives with a paying design partner and a compliance answer is underwriting with that filter instead of against it.

The pattern already runs elsewhere in the portfolio. Alphajuri applies it to Brazilian judicial assets, precatórios and claims. WIR, with AXA, applies it to async insurance pricing and risk scoring. Neither publishes metrics, so take the pattern rather than an outcome. Three moves, in a fixed order.

• Copilot. Sell an underwriting or claims copilot to a mid-market lender, insurer or MGA with no in-house data science team. Revenue from day one, not a data-collection pretext dressed as a product.
• Data. Every decision the copilot supports generates a labelled outcome. Over 18 to 36 months that becomes a loss and behaviour dataset on a specific book, in a specific segment, that no incumbent and no foreign entrant holds.
• Fund. That dataset becomes the underwriting edge for a downstream credit or insurance vehicle. The software business proves the model. The capital vehicle monetises the edge at a completely different margin structure.

BACEN and SUSEP, explainability, and incumbent balance sheets

Explainability in Brazil is a legal requirement, not a design preference. Article 20 of the LGPD gives any data subject the right to request review of decisions taken solely on automated processing that affect their interests, expressly including credit profiling, and obliges the controller to supply clear and adequate information on the criteria and procedures used. A model that cannot produce that record is not deployable in a credit or claims decision. Not risky. Not deployable.

The AI-specific rules are coming and are not settled, which is a risk in both directions. PL 2338/2023, the Marco Legal da IA, was approved unanimously by the Senate on 10 December 2024 and remains before the Chamber of Deputies. It follows the EU AI Act structure with risk tiers, rights to explanation and contestation, a national governance system linked to the ANPD, and sanctions of up to R$ 50 million per infraction. Banco Central placed AI on its 2025 to 2026 regulatory priorities agenda as studies on the risks of AI use in the financial system, aiming at guidelines rather than immediate binding rules. SUSEP set its 2026 regulation plan through Resolução SUSEP nº 72/2025, with explainability of automated underwriting and claims decisions a central theme.

So there is no safe harbour yet. Build to the strictest plausible standard now, because retrofitting a decision log into a shipped model costs more than designing one in, and the retrofit always lands in the same week the first regulator asks.

The incumbent is not a sleepy target either. Itaú Unibanco reported roughly 150 generative AI solutions already in production and more than 750 projects in development by early 2026. Its ia.i assistant reached 300,000 users by July 2026 against a stated target of the full base of more than 60 million clients by December. Bradesco's Bridge platform, built on Azure OpenAI, reports an 83 percent external resolution rate across roughly 74 million customers, a technology cost reduction above 30 percent, and solution launches up to 10 times faster.

So do not compete with the five largest Brazilian banks on general-purpose retail assistants. They have the teams, the balance sheets and the distribution. Compete where they are structurally slow, which is the mid-market institution they do not serve and the vertical workflow too small to make their roadmap.

Then price the calendar honestly. BFSI procurement in Brazil runs in quarters, not weeks, and passes through information security, compliance, legal and often a regulator-facing risk committee. Budget 6 to 12 months from first meeting to signed pilot. The compliance answer has to exist before the product demo, not after it.

How Avante would approach it

Avante Ventures is a venture studio building AI-native companies in Brazil and Latin America. Not an accelerator, not an incubator, not a fund. The distinction earns its keep in BFSI more than in most categories, because the scarce input here is neither capital nor engineering. It is a person who has already owned the decision you intend to automate, inside a Brazilian institution, through a full rate cycle.

Inference cost for GPT-3.5-level capability fell more than 280-fold between November 2022 and October 2024, from USD 20.00 to USD 0.07 per million tokens, per the Stanford HAI 2025 AI Index. A vertical BFSI product that needed a Series A to fund inference in 2022 now reaches first revenue on a seed budget. That is the concrete basis for deploying AI without a Series A. The binding constraint moved from compute to distribution and compliance, which is exactly the pairing a studio staffs on day one and a traditional fund has to sit and wait for. More on how we assemble it is on why Avante operates as a venture studio.

Avante deploys $500K-$1.5M per venture across pre-seed and retains co-founder economics, launching 3-4 ventures per year. Solving the company plumbing once routes roughly $300K-$500K of effective capital per venture into product and traction rather than overhead, and a studio venture typically launches 6-9 months ahead of a comparably funded standalone team. Operating partners stay engaged through the first revenue milestone. In BFSI that head start compounds harder than anywhere else, because the sales cycle is measured in quarters and the calendar is the scarce input.

The substrate is finished. The regulation is half-written. The incumbents are awake and spending R$ 47.8 billion a year on technology. What is still unclaimed is the mid-market decision nobody has automated with an audit trail attached, and in Brazil the clock on that started the day the first Open Finance consent was granted. Here is how the six-stage system of Research, Partner, Build, Traction, Revenue, Compound runs against it.

• Research. Pick the decision, not the sector. Underwriting for a mid-market lender is a company. Generative AI for banking is a category page.
• Partner. Recruit the domain operator with 10+ years of Brazilian-market scar tissue who has already owned that decision inside an institution. This is the binding constraint, not engineering.
• Build. Ship the decision log and the compliance posture in version one, because in BFSI the audit trail is the product surface.
• Traction, Revenue and Compound. One paying design partner before the second engineer. Then a contract that survives a compliance review, which is the test that actually matters. Then the accumulated outcome data turned into the underwriting edge for a downstream vehicle. Copilot, then data, then fund.

Frequently asked questions

How large is the Brazil generative AI in BFSI market?

No research house publishes a defensible Brazil-specific figure for generative AI in BFSI, and estimates for the same base year diverge by more than 2x. IMARC Group sizes the whole Brazil generative AI market at USD 371.2 million in 2025, reaching USD 1,481.5 million by 2034 at a 16.63 percent CAGR. The better anchor is observed spend. FEBRABAN and Deloitte put the Brazilian bank technology budget at R$ 47.8 billion in 2025, with AI and analytics budgets up an estimated 61 percent.

Why is Brazil a better place to build generative AI for banking than the United States?

Because permissioned financial data is portable by regulation rather than by partnership. Open Finance Brasil reported 208.79 million active consents as of 31 July 2026 and more than 2.3 billion successful API communications per week, and participation is mandatory for large institutions. Pix adds near universal payment coverage, with 79.8 billion transactions worth R$ 35.36 trillion in 2025 and more than 170 million users.

Can generative AI be used for credit decisions in Brazil?

Yes, with conditions that are legal rather than technical. Article 20 of the LGPD grants a right to review of decisions taken solely on automated processing, expressly including credit profiling, and requires clear information on the criteria used. PL 2338/2023 passed the Senate in December 2024 and would add EU-style risk-tier obligations with sanctions of up to R$ 50 million per infraction. Any deployable model has to emit an auditable reasoning record.

Why is the Brazil AI insurance market a growth story rather than a mature one?

Brazilian insurance penetration was 3.3 percent of GDP in 2024 against an OECD average of 6.2 percent, with both measured as premiums over GDP. Brazil sits at roughly half the developed-market level while the sector keeps expanding, and CNseg projects R$ 808 billion in 2026 revenue, up 5.7 percent. Low penetration on a large and growing base is a growth argument.

Are incumbent Brazilian banks already building this?

Yes, at serious scale. Itaú Unibanco reported roughly 150 generative AI solutions in production and more than 750 in development by early 2026, and Bradesco's Bridge platform reports an 83 percent external resolution rate with technology costs down more than 30 percent. The opening is not general retail assistants. It is the mid-market institution the top five do not serve and the vertical workflow too small for their roadmap.