Between 70% and 85% of AI projects in financial services fail to reach production or deliver measurable business value, according to research from Gartner, FinTellect, and the RAND Corporation. The failure rate is consistent across markets. In Latin America, the cost of that failure is uniquely compounded: Pix volumes grow monthly, the fraud surface expands with each new transaction on an irrevocable rail, and MED 2.0 enforcement has been active since May 2026. Every month an institution spends on a failing AI programme is a month the fraud gap widens and the competitive disadvantage against digital-first banks — Nubank, Mercado Pago, PicPay — narrows the territory available to incumbents.
The reason programmes keep failing is the same in LATAM as everywhere else. Institutions are targeting the wrong decisions first, and when they do target the right ones, they are running programmes that are structurally set up not to deliver regardless of what they are pointed at.
Volume and value are not the same thing — and in LATAM, the highest-value decisions are specific
The standard misprioritisation applies in LATAM as in every other market: teams build AI cases around high-volume, lower-consequence decisions because the automation saving is legible, the sponsor is easy to identify, and success can be measured in straightforward cost terms. The decisions that attract AI investment first are account administration, STP routing, and low-value authorisations. These are not wrong to automate. They are not where the value sits.
In LATAM, the two decisions with the clearest and most immediate value at stake are Pix fraud prevention and thin-file credit origination, and the reasons are specific to the region. Pix fraud prevention is not merely high-value — it is architecturally mandatory. The irrevocable settlement model means that the window for prevention is measured in milliseconds before payment initiation. Fraud identified after the payment settles is fraud that has already succeeded. The only decision system that can prevent rather than detect Pix fraud is one that operates inline, in real-time, at the point of initiation. MED 2.0 requires multi-hop transaction tracing that rules-based systems cannot provide. This is not a matter of degree — it is a categorical distinction between what AI can do and what rules cannot.
Thin-file credit origination is the other LATAM-specific priority, and it is the one most commonly underestimated. The instinct is to treat it as a financial inclusion initiative — a socially valuable but commercially marginal effort to serve underserved populations. That framing is wrong. The 48% labour informality rate across LATAM means that roughly half the workforce is commercially invisible to bureau-based scoring. These customers are not absent from the banking system — Pix’s penetration makes clear that tens of millions of them are actively transacting. They are absent from credit models, which means they are absent from credit revenue. AI models trained on alternative signals — Pix transaction frequency and regularity, mobile payment patterns, utility payment history — can assess creditworthiness for this population profitably. The question is not whether to serve them. The question is whether AI infrastructure will be in place to do so before digital-first competitors build the outcome data advantage that will make the gap structural.
Regulatory urgency in LATAM functions differently from other regions. There is no single deadline like the EU AI Act or CPS 230. The urgency is transactional: with every quarter of inaction, the fraud surface grows, MED 2.0 non-compliance accumulates, and competitor models accumulate outcome data that the institution does not. The pressure is continuous rather than threshold-based, which makes it easier to defer — and makes the consequences of deferral compound more quietly but no less materially.
The three reasons AI programmes fail — and what is specific about failing in Brazil
For institutions that correctly identify the right decisions and commit to targeting them, the failure patterns are consistent across every market. They break into three categories in roughly the same proportions every time.
Governance failure accounts for around 42% of failed initiatives. The pattern: technology selected before the problem was defined, the proof of concept proved the model works, nobody was assigned to act on it. No named decision, no named sponsor, no measurable outcome, no production funding pathway at the start. In the Brazilian context, governance failure has a specific dimension: a model built for Pix fraud prevention that reaches production without a clear decision architecture — specifically, without the inline scoring infrastructure to operate before payment initiation — is not a fraud prevention system. It is a fraud detection system. The difference is not marginal. The difference is whether the bank prevents the loss or detects it after it has occurred and then tries to recover funds that MED 2.0 may or may not trace.
Sequencing failure accounts for around 31%. Data was not ready. Nobody admitted it until the proof of value was already running. Teams proceeded with inadequate labelled outcome data and produced models too weak to deploy. In LATAM, the data readiness question has a specific texture for thin-file credit models: the relevant data — Pix transaction history, mobile payment patterns, alternative signals — may exist but not be aggregated, not be labelled with outcome data, and not be accessible through the pipelines required for model training. Building the data infrastructure to support alternative-signal credit models requires lead time that is commonly discovered only after the modelling work has begun.
Measurement failure accounts for the remaining 27%. The model achieved its technical targets but nobody defined what a good business outcome looked like in dollar terms. In the Pix fraud context, measurement failure has a direct cost: a model that achieves high accuracy in offline evaluation but does not have an economic baseline — what was the fraud loss rate before, and what is it now — cannot build the production funding case that enterprise rollout requires. More precisely, it cannot answer the question MED 2.0 enforcement will eventually ask: was this institution’s fraud detection capability adequate, and do you have the evidence to demonstrate it?
| Failure mode | Share of failed initiatives | LATAM-specific dimension |
|---|---|---|
| Governance failure | ~42% | Model in production without inline scoring architecture is a detection system, not a prevention system. The distinction is the entire value of the programme. |
| Sequencing failure | ~31% | Alternative-signal data pipelines for thin-file credit require lead time typically discovered only after modelling starts. LGPD localisation adds infrastructure requirements. |
| Measurement failure | ~27% | No economic baseline means no MED 2.0 adequacy evidence. Enforcement produces the question; measurement failure means no answer. |
Source: LATAM Banking Practice analysis, consistent with findings across North America, EMEA, APAC, and ANZ markets.
The structural implication holds everywhere and is particularly acute in Brazil: investment in model sophistication generates diminishing returns if governance, sequencing, and measurement conditions are not in place. The most capable fraud model that operates off the critical path — scoring after initiation rather than before — delivers exactly nothing in loss prevention terms, regardless of its AUC score. Programme design is not downstream of technology selection. It determines whether the technology investment produces any return at all.
What this means for sequencing in LATAM
The correct sequencing in Brazil almost always produces the same first answer: real-time Pix fraud scoring, inline at the point of initiation. It has the highest volume of any decision category, the highest financial consequence per error on an irrevocable rail, data readiness that Pix transaction history provides, and live regulatory urgency from MED 2.0. No other decision category combines all four criteria as strongly.
Thin-file credit origination is the second priority for most institutions — the data readiness question requires assessment, but the commercial opportunity is large and the competitive pressure from digital-first banks is immediate. AML transaction monitoring is the third priority, with regulatory expectations from the BCB’s Resolution 6 on inter-institutional fraud intelligence sharing adding governance urgency.
The value of making this sequencing explicit and systematic — rather than intuitive — is not the output. It is the process: a structured cross-functional conversation that surfaces data readiness gaps before commitment is made, names the sponsor who will own production outcomes, and establishes the economic baseline against which the programme will be measured. These are precisely the conditions whose absence drives the majority of LATAM AI programmes to fail.
Part 2 of 3.
Sources
Gartner. AI Project Failure Rates. Multiple editions, 2023–2025. FinTellect AI. Why 80% of AI Projects in Finance Fail. 2024. RAND Corporation. AI Project Success and Failure Rates. Referenced in industry literature. International Labour Organization. Labour Informality in Latin America. Mid-2023 data. Banco Central do Brasil. MED 2.0 enforcement framework. Mandatory February 2026, penalties from May 2026. LATAM Banking Practice. Decision Portfolio and Failure Mode Analysis. Internal analysis, consistent across North America, EMEA, APAC, and ANZ markets.
