Technology & AI
Why We Build Everything on Anthropic Claude
Why Innova Black chose Anthropic Claude as the backbone for fintech compliance in Mexico.
At Innova Black we made a decision that defined our future as a financial-technology firm: building our entire artificial-intelligence infrastructure on Anthropic Claude. It was not impulsive nor a bet on fashion. It was the result of months of rigorous evaluation, production tests with real compliance data and a conviction that reinforced with every iteration: when you work with Mexican financial regulation, you cannot afford an AI that invents, hallucinates or gives you a creative answer when what you need is absolute regulatory precision.
This article is our technology manifesto. We explain why we chose Anthropic, how we integrate Claude into every module of our DTX platform, what results we have achieved and where we are heading. If you direct a SOFOM, lead compliance at a fintech or simply want to understand how frontier AI applies to the regulated financial sector in Mexico, this concerns you directly.
Because the question is no longer whether AI will transform financial compliance. The question is which AI you will build your operation on. And that choice has consequences that go far beyond technology.
The problem: Generic AI does not work in financial regulation
Before talking about solutions, the problem must be understood clearly. Mexican financial regulation is not a domain where you can tolerate approximations. The CNBV AML/CFT General Provisions are not suggestions: they are legal mandates whose non-compliance generates fines exceeding MXN 11 million per infraction. A single interpretation error in a KYC requirement, a miscalculated deadline for an Unusual Operations Report, a wrong risk classification in a file — each of these scenarios can trigger a sanctioning procedure.
We tested generic AI models in this context. The results were alarming. Models that performed brilliantly to generate code, summarize text or answer general questions failed in dangerous ways when faced with Mexican financial regulation. Errors were not innocent: they were articles of law that did not exist, invented deadlines, requirements the model arbitrarily combined from different provisions, and citations to repealed circulars presented with full confidence.
This phenomenon — known as “hallucination” — is structurally unacceptable in regulatory compliance. A SOFOM cannot file a report based on a deadline the model invented. A compliance officer cannot rely on a list of requirements an AI fabricated. A general manager cannot make decisions based on a regulatory analysis that mixes valid provisions with phantom provisions. The legal, financial and reputational cost of an AI hallucination in this context is, literally, the future of the institution.
Hallucinations are not a bug an update fixes: they are a structural feature of generative models that prioritize fluency over factuality. Models like ChatGPT have inherent architectures that fill information gaps with statistically probable content, regardless of whether that content corresponds to verifiable reality. For domains where being wrong has costs in millions, this is a fundamental disqualification.
Why Anthropic Claude
Anthropic was founded in 2021 by Dario Amodei and Daniela Amodei, alongside other former OpenAI researchers who departed because of philosophical differences about AI safety. The original premise of Anthropic is direct: AI must be developed with safety as the central priority, not as an afterthought once a product is launched.
This decision is not academic. It translates into concrete technical practices that radically change the behavior of the resulting models. The “Constitutional AI” methodology Anthropic uses to train Claude implements a multi-layered system of principles that guide the model toward safer, more honest and more useful responses. Among these principles, the disposition to recognize the limits of one’s own knowledge stands out — what in technical terms is called calibrated uncertainty.
Claude’s calibrated uncertainty is the single most important feature for our use case. When we ask Claude about a specific provision of Mexican financial regulation and the model does not have sufficient confidence in the answer, Claude tells us. It does not invent. It does not improvise. It does not generate a confident-sounding response to hide the gap. It simply tells us it does not have enough information to answer with the required certainty, and suggests verifying directly in the original source.
This is exactly the behavior we need for our domain. A SOFOM building compliance reports cannot accept an AI that improvises when uncertain. It can only accept an AI that, when it does not know, says so — so the human professional can intervene and provide the correct information.
Constitutional AI: AI guided by principles
To understand why Claude behaves differently in regulatory contexts, the methodology Anthropic uses to train it must be understood. The “Constitutional AI” approach implements a set of principles — a “constitution” — guiding the model’s behavior during training. These principles include explicit values such as honesty, harm avoidance, recognition of uncertainty and respect for the well-being of the user.
This is not philosophical decoration. The constitution is implemented through specific reinforcement-learning techniques that systematically reward responses adhering to the principles and penalize those violating them. The cumulative result over thousands of training iterations is a model that has internalized these values to the point of expressing them naturally in its responses.
When you ask Claude to interpret a provision of LFPIORPI, the model not only retrieves information about the regulation. It also applies its principles of honesty (it does not say more than it knows), harm avoidance (it does not provide advice that could lead the user to non-compliance) and recognition of uncertainty (it specifies when the answer requires expert verification). This behavior is exactly what we need for Mexican financial regulation.
How we integrate Claude into our DTX platform
Our DTX platform is not a wrapper that calls Claude’s API and presents the result to the user. It is a sophisticated integration that combines Claude with specialized regulatory databases, real-time validation logic and human review workflows for critical decisions.
Specialized regulatory ingestion. We maintain a structured database of Mexican financial regulation that includes all CNBV General Provisions, LFPIORPI, LGOAAC, the Banking Law (LIC), the Fintech Law (LRITF) and their reforms, official UIF criteria and relevant SCJN theses. This database is updated automatically when changes are published in the Official Gazette of the Federation.
Context retrieval before generation. When a user asks a regulatory question, the system does not pass the question directly to Claude. First it retrieves the relevant context from our regulatory database, including the specific articles applying to the case. Then it passes that context to Claude with the original question, allowing the model to base its answer on actual provisions and not on its parametric memory.
Mandatory citation of sources. Every answer Claude generates in our platform must include explicit citation of the regulatory sources it references. If the model cannot cite a specific source, the answer is flagged as “unverified” and presented to the user with a warning recommending consultation with a specialist.
Automatic validation against ground truth. For critical operations such as Unusual Operations Reports, the system implements a second validation layer that compares Claude’s response against decision rules codified by our compliance experts. If there is a discrepancy between Claude’s response and the codified rules, the system escalates to human review before executing the action.
Workflows with human-in-the-loop. For sensitive decisions — generation of regulatory reports, risk classifications, validations of high-value operations — the system always requires explicit human approval. Claude generates the recommendation; the qualified human professional validates and approves. This approach combines the speed and consistency of AI with the legal accountability and judgment of the responsible professional.
Specific use cases: Claude in production
Generation of Unusual Operations Reports (ROU and ROIP). When the AML system detects an operation that meets unusual criteria, Claude generates a draft regulatory report including narrative of the operation, technical justification of the unusual qualification, list of relevant elements and references to the specific LFPIORPI articles. The compliance officer reviews, edits if necessary and approves. Generation time went from 2 to 3 hours per report to 15 to 20 minutes.
KYC analysis of complex corporate structures. When a SOFOM evaluates a corporate client with complex structure (holdings, multiple subsidiaries, foreign investors), Claude analyzes the structure, identifies beneficial owners, evaluates jurisdictional risk and generates documented recommendations. What used to take a junior analyst 4 to 6 hours, Claude does in 10 minutes with quality equivalent to senior analyst.
Interpretation of regulatory queries. Compliance officers can ask specific questions about Mexican financial regulation — “what are the documentation requirements for a high-risk PEP client”, “what is the deadline to file an ROIP”, “how is enhanced risk assessment applied to foreign clients” — and receive precise answers with citation of articles and applicable provisions.
Automated compliance audit. Claude analyzes compliance documents — manuals, policies, generated reports, training records — and identifies inconsistencies, gaps or non-conformities with regulation. This allows internal audit teams to focus their attention where it is really needed, instead of manually reviewing thousands of documents.
Drafting and review of compliance policies. When a SOFOM needs to update its AML manuals, Claude generates initial drafts based on current regulation, specific industry segment and institution-specific characteristics. The result is a manual that is not generic, but tailored to the institution and aligned with the regulation.
The results: data on Claude vs alternatives
After more than 18 months of operation with Claude in production, we have accumulated quantitative data that supports our choice.
Factual accuracy in regulatory questions. In standardized tests with 500 specific questions about Mexican financial regulation, Claude achieved 94% factual accuracy with citation of correct sources. Generic models we evaluated as comparison ranged between 67% and 78% accuracy, with frequent hallucinations of articles and inexistent provisions.
Calibrated uncertainty. Of Claude’s responses identifying themselves as “uncertain” or “requires expert verification”, 89% genuinely corresponded to cases where the answer required specialized verification. Of generic models’ responses with similar self-identification, only 34% genuinely corresponded to uncertain cases — the rest were responses where the model could have answered with confidence but did not.
Document analysis quality. In a blind evaluation with 30 senior compliance experts, Claude’s reports on KYC of complex corporate structures were rated, on average, 8.7/10 on quality and completeness. Generic models’ reports were rated 6.2/10. Human senior analyst reports were rated 9.1/10 — meaning Claude approaches but does not exceed expert human performance, as expected.
Cost efficiency. Despite Claude being more expensive per token than some generic alternatives, the total cost per task ended up being significantly lower because of the reduced rework. We do not have to fix hallucinations, validate fabricated citations or correct invented procedures. The professional time saved more than compensates the higher unit cost of the model.
Looking ahead: where we are going with Claude
Our trajectory with Anthropic is not limited to using Claude as it exists today. We are actively working on three frontiers that will define our value proposition over the next 24 months.
Increasingly specialized fine-tuning. We are developing fine-tuned versions of Claude specifically trained on Mexican financial regulation, our compliance experts’ criteria and patterns identified in CNBV inspections. This allows us to deliver responses even more aligned with the practical reality of regulated entities in Mexico.
Autonomous regulatory agents. Beyond responding to questions, we are building agents that proactively execute compliance tasks: continuous transactional monitoring, anomaly detection, automatic generation of preliminary reports for human review, regulatory updates in manuals when relevant changes are published. The vision is that the institution’s compliance team focuses on judgment and strategic decision, not repetitive operational tasks.
Auditable explainability. A unique feature we are developing is the system’s ability to explain in detail any decision, recommendation or alert it generates. CNBV in inspection can request why a specific operation was classified as unusual, why a client was assigned a certain risk level, why a specific report was filed — and our platform must be able to provide a complete, technical and verifiable answer.
What this means for your institution
If you direct a financial institution in Mexico — SOFOM, SOFIPO, IFPE, factor, leasing firm — you face an inevitable strategic decision: which technological infrastructure are you going to build your compliance operation on? The decision has implications you may not see today but you will live with for years.
A SOFOM that today operates compliance manually or with limited tools will soon face an increasingly hostile environment. The FATF mutual evaluation in 2026 will intensify CNBV’s expectations on Mexican financial entities. Authorized fintechs are operationalizing AI compliance now and creating a gap traditional SOFOMs will struggle to close. The bar to operate competitively goes up every quarter.
Choosing the right technological infrastructure is not a luxury. It is a strategic necessity that determines whether your institution will be among the ones that scale or among the ones that get displaced. And in that choice, the AI underpinning your compliance operation is one of the most important decisions you will make.
For us at Innova Black, that decision is taken: Anthropic Claude. Not as a fashionable bet, but as a structural choice grounded in technical analysis, production validation and conviction about what financial regulation requires. We share this analysis transparently because we believe regulated Mexican fintech needs more rigorous conversations about technological infrastructure — not more marketing buzz about generic AI.
If your institution is evaluating how to incorporate AI into your compliance operation, we are happy to share our experience. The technological landscape evolves quickly and the decisions you take today define the trajectory of your institution for the next five years.