This case study from Adyen, a global payments platform company, discusses their approach to creating value through AI Applied Research Engineering. Published in June 2025, the article by Andreu Mora, SVP and Global Head of Engineering Data at Adyen, appears to explore how the company leverages AI research and engineering practices to enhance their payment processing and risk management capabilities. While the provided text is primarily navigational content from a webpage rather than the full article, it indicates Adyen's strategic focus on applying AI research methodologies within their engineering organization to unlock business value in the fintech domain.
This case study represents Adyen’s approach to implementing AI Applied Research Engineering within their global payments infrastructure. Adyen is a major player in the financial technology sector, providing payment processing, risk management, authentication, and issuing services across multiple markets globally. The article, authored by Andreu Mora who serves as SVP and Global Head of Engineering Data, was published on June 5th, 2025, positioning it as recent insights into how a leading fintech company is approaching AI implementation at scale.
It’s important to note that the provided text consists primarily of webpage navigation elements, footer content, and metadata rather than the full substantive content of the article itself. This limitation means we can only infer the general themes and directions of Adyen’s AI implementation strategy rather than examining specific technical implementations, architectural decisions, or quantified results. The title “Unlocking value through AI Applied Research Engineering” suggests a focus on bridging research and production engineering, which is a critical aspect of successful LLMOps.
Adyen operates in the highly regulated and performance-critical financial services sector, specifically in payment processing. This context is crucial for understanding their LLMOps requirements. Payment platforms face unique challenges including real-time processing requirements (transactions must be approved or declined in milliseconds), stringent regulatory compliance across multiple jurisdictions, fraud detection and prevention at massive scale, and the need for extremely high reliability and availability given the financial stakes involved.
The company’s global footprint, evidenced by their localized presence across Australia, Brazil, China, various European markets, Japan, Mexico, the UAE, Singapore, and the UK, indicates they must handle diverse payment methods, currencies, languages, and regulatory frameworks. This international scope suggests that any AI or LLM implementations would need to be multilingual, culturally aware, and adaptable to different regulatory environments.
The concept of “Applied Research Engineering” as framed in the title suggests a hybrid approach that sits between pure research and traditional software engineering. In the context of LLMOps, this philosophy is particularly relevant because it addresses one of the industry’s most persistent challenges: the “research-to-production gap.” Many organizations struggle to translate AI research breakthroughs and experimental models into robust, scalable production systems that deliver business value.
The applied research engineering approach typically involves several key elements that are critical for successful LLMOps implementation. First, there’s the creation of teams that combine research expertise with production engineering skills, allowing for rapid prototyping of novel AI approaches while maintaining production-grade quality standards. Second, this approach emphasizes iterative development cycles where research insights quickly inform product features, and production feedback loops back to guide research priorities. Third, there’s a focus on building reusable infrastructure and frameworks that can accelerate the path from research concept to deployed feature.
In the payments and fintech context, applied research engineering for LLMs might address several key use cases. These could include fraud detection and risk assessment, where LLMs could analyze transaction patterns, merchant descriptions, and customer behavior in natural language to identify suspicious activities. Customer support automation is another natural application, with conversational AI handling payment inquiries, dispute resolution, and merchant onboarding questions across multiple languages. Merchant categorization and understanding could leverage LLMs to better understand business descriptions, product catalogs, and transaction descriptions for more accurate merchant category coding and risk profiling. Regulatory compliance and document processing might employ LLMs to interpret regulatory documents, automate compliance reporting, and extract information from financial documents. Additionally, internal tooling and developer experience could be enhanced through AI-assisted coding tools, documentation generation, and intelligent search across Adyen’s technical knowledge base.
Operating LLMs in production within the financial services sector presents unique challenges that would shape Adyen’s approach to LLMOps. Regulatory compliance and auditability are paramount - all AI decisions, especially those affecting transaction approval or risk scoring, must be explainable and auditable for regulators. This requirement may favor approaches that combine LLMs with more traditional, interpretable models, or necessitate sophisticated explainability layers on top of LLM outputs.
Data privacy and security are equally critical given that payment data is highly sensitive and subject to strict data protection regulations like GDPR and PCI-DSS. Any LLM implementation must ensure customer and merchant data is properly anonymized, encrypted, and accessed only by authorized systems. The use of LLMs likely requires careful data governance, potentially including techniques like differential privacy, federated learning, or working exclusively with carefully curated and sanitized datasets.
Latency and performance requirements in payment processing are extremely stringent. While some use cases like customer support or merchant onboarding can tolerate response times in the seconds range, transaction processing decisions must happen in milliseconds. This constraint means that LLMs, which can be computationally expensive, would need to be carefully optimized for inference speed or reserved for asynchronous processes. The LLMOps infrastructure would need to support multiple deployment patterns optimized for different latency requirements.
Accuracy and reliability standards in fintech are notably higher than in many other domains. A chatbot error in e-commerce might be annoying; an AI error in payment processing could result in financial losses, regulatory violations, or damaged merchant relationships. This drives the need for extensive testing frameworks, staged rollouts, continuous monitoring, and robust fallback mechanisms that are core to mature LLMOps practices.
While the provided text doesn’t detail specific infrastructure choices, we can infer certain requirements based on Adyen’s operational profile. As a global platform processing payments across multiple regions, Adyen would need geographically distributed LLM inference infrastructure to minimize latency and comply with data residency requirements. This likely involves deploying models close to where data originates and transactions are processed.
The company’s focus on reliability would necessitate robust LLMOps infrastructure including model versioning and rollback capabilities, A/B testing frameworks to safely compare model versions in production, comprehensive monitoring and observability for model performance and business metrics, automated testing pipelines for both technical performance and business outcome validation, and disaster recovery and failover mechanisms to ensure continuous operation.
Given the scale of operations (major payment processors handle millions of transactions daily), the infrastructure must support high-throughput inference, potentially requiring specialized hardware like GPUs or custom inference accelerators, sophisticated load balancing and auto-scaling, and efficient model serving architectures that can handle variable load patterns throughout the day and across different geographic regions.
Andreu Mora’s role as Global Head of Engineering Data suggests that data infrastructure and engineering are central to Adyen’s AI strategy. Effective LLMOps requires robust data pipelines that can collect, clean, label, and version training data at scale. In the payments domain, this involves aggregating data from numerous sources including transaction logs, merchant profiles, customer interaction histories, fraud reports and dispute outcomes, and external data sources for context enrichment.
The data engineering challenges in implementing LLMs for payments would include maintaining data quality and consistency across global operations, implementing privacy-preserving data processing techniques, creating efficient feature stores that can serve both training and real-time inference, and establishing processes for continuous data collection to support model retraining and improvement.
Model development in this context would need to balance innovation with risk management. Applied research engineering teams would likely employ techniques such as transfer learning from large foundation models, fine-tuning on domain-specific payment and financial data, ensemble approaches combining LLMs with traditional ML models, and rigorous offline evaluation before any production deployment.
Financial services companies must implement strong governance around AI systems. For LLMOps, this includes model risk management frameworks that assess and monitor risks associated with LLM deployment, bias detection and mitigation to ensure fair treatment across different customer and merchant populations, compliance validation ensuring outputs meet regulatory requirements, and human-in-the-loop systems for high-stakes decisions.
Adyen would need clear ownership and accountability structures defining who is responsible for model performance, compliance, and incident response. The applied research engineering approach helps here by creating teams that understand both the technical capabilities and limitations of LLMs and the business and regulatory context in which they operate.
Production LLMOps requires comprehensive evaluation and monitoring strategies. For Adyen, this would likely include both technical metrics (latency, throughput, error rates, model confidence distributions) and business metrics (fraud detection rate, false positive rates for risk scoring, customer satisfaction with AI interactions, operational cost savings). The monitoring must be real-time for critical applications, with alerting systems that can detect model drift, performance degradation, or anomalous outputs that might indicate issues.
Continuous evaluation would involve comparing model predictions against ground truth outcomes (e.g., did transactions flagged as risky actually turn out to be fraudulent), A/B testing to measure the incremental value of new model versions, and regular audits to ensure compliance with internal policies and external regulations.
While the article title emphasizes “unlocking value,” a balanced assessment must acknowledge the challenges inherent in implementing LLMs in production for financial services. The complexity of integration with existing systems is significant - Adyen operates a sophisticated payment platform, and integrating LLMs requires careful orchestration with existing fraud detection systems, transaction processing pipelines, and customer relationship management tools.
The cost-benefit analysis must be rigorous. LLMs can be expensive to train, deploy, and operate at scale. Organizations must carefully evaluate whether the value generated (improved fraud detection, better customer experience, operational efficiency) justifies the investment in infrastructure, specialized talent, and ongoing operations.
Talent acquisition and retention presents another challenge. Building applied research engineering teams requires professionals with both deep AI expertise and understanding of production engineering and the payments domain - a rare combination. The “brain drain” from financial services to pure tech companies can make this especially challenging.
There are also inherent technical limitations of LLMs that must be managed. These models can hallucinate plausible-sounding but incorrect information, which is unacceptable in financial contexts. They can be brittle when encountering out-of-distribution inputs. They require large amounts of data and compute. Managing these limitations requires careful system design, extensive testing, and often hybrid approaches that combine LLMs with more traditional, deterministic systems.
The applied research engineering model implies certain organizational structures and cultural attributes. Successfully implementing this approach requires breaking down silos between research and engineering teams, creating career paths that value both research excellence and production impact, and fostering a culture of experimentation balanced with operational discipline.
For LLMOps specifically, this means creating environments where data scientists and ML engineers can rapidly experiment with new models and approaches, but within guardrails that ensure production stability and compliance. It requires tooling and platforms that make it easy to go from experimental notebook to production-grade service, and feedback mechanisms that help research teams understand how their work performs in real-world conditions.
The fact that Adyen is publicly discussing their AI Applied Research Engineering approach in 2025 reflects the maturation of AI implementation in financial services. It suggests a shift from viewing AI as experimental or nice-to-have to treating it as core infrastructure that requires dedicated engineering investment and specialized operational practices.
For the broader industry, Adyen’s approach may signal several trends in LLMOps. There’s movement toward hybrid teams that combine research and engineering skills rather than keeping these functions separate. We’re seeing increasing emphasis on applied research that’s directly connected to product and business outcomes rather than pure research. There’s growing recognition that LLMOps is a distinct discipline requiring specialized tools, platforms, and practices. And there’s an understanding that successful AI implementation requires both technical excellence and deep domain expertise, especially in regulated industries.
This case study, based on limited publicly available information, represents Adyen’s strategic direction in applying AI and potentially LLMs to create value in payment processing and financial services. While the specific technical implementations, architectural decisions, and quantified results are not detailed in the provided text, the framing around “Applied Research Engineering” suggests a mature approach to bridging research innovation and production deployment.
The payments industry presents both opportunities and challenges for LLMOps implementation. The opportunities include massive scale of data for training, clear business value in fraud prevention and customer experience, and potential for significant operational efficiency gains. The challenges include stringent regulatory requirements, extreme performance and reliability standards, and the high stakes of getting AI decisions wrong in financial contexts.
It’s important to approach this case study with appropriate skepticism given that it originates from corporate communications materials likely designed to position Adyen as an AI leader. Without access to detailed technical implementations, independent verification of results, or peer review of methodologies, we cannot fully assess the effectiveness of their approach. However, the strategic emphasis on applied research engineering, the senior leadership attention reflected in authorship by a SVP-level executive, and the public commitment to this direction all suggest serious organizational investment in mature AI and LLMOps practices.
For organizations looking to implement similar approaches, key takeaways would include the importance of bridging research and production through dedicated applied research engineering teams, the need for domain-specific considerations in LLMOps (especially in regulated industries), the value of strong data engineering foundations for any AI initiative, and the requirement for comprehensive governance, evaluation, and monitoring frameworks when deploying LLMs in high-stakes production environments.
Block faced the challenge of maintaining system resilience at scale as engineering teams shipped locally rational but globally corrosive features that eroded overall architecture. They developed "Builderbot," an agentic code review system that acts as a vigilant guardian rather than a passive assistant, continuously observing, learning, and steering changes to align with their organizational "world model." The solution shifts protection left in the development lifecycle, uses standardized CLI contracts (Just) for local development, implements progressive context disclosure through AGENTS.md files and Code Review Checks, and leverages Agent Skills for dynamic context loading. The result is a protector system that enables velocity with confidence, catching issues pre-push, reducing burden on human reviewers, and ensuring architectural alignment across the entire organization.
Ramp, a finance automation platform serving over 50,000 customers, built a comprehensive suite of AI agents to automate manual financial workflows including expense policy enforcement, accounting classification, and invoice processing. The company evolved from building hundreds of isolated agents to consolidating around a single agent framework with thousands of skills, unified through a conversational interface called Omnichat. Their Policy Agent product, which uses LLMs to interpret and enforce expense policies written in natural language, demonstrates significant production deployment challenges and solutions including iterative development starting with simple use cases, extensive evaluation frameworks, human-in-the-loop labeling sessions, and careful context engineering. Additionally, Ramp built an internal coding agent called Ramp Inspect that now accounts for over 50% of production PRs merged weekly, illustrating how AI infrastructure investments enable broader organizational productivity gains.
Brex, a financial services company, faced a significant challenge with customer onboarding that took days due to manual Know Your Customer (KYC) and underwriting processes that relied on implicit heuristics and manual judgment. To solve this, they rebuilt their entire onboarding system as an AI-native, multi-agent architecture where specialized agents collaborate through structured reasoning to handle verification, fraud detection, document processing, and underwriting decisions. The results were dramatic: they moved from 0% to 40% auto-approval of card applications in weeks, reduced manual identity reviews by 70% through specialized fuzzy-matching agents, achieved 85% reduction in business address requests for information (RFIs), and enabled most eligible businesses to onboard in minutes rather than days while maintaining or improving accuracy and creating full auditability trails for every decision.
